Relational Inductive Biases, Deep Learning, and Graph Networks

Relational inductive biases, deep learning, and graph networks
Peter W. Battaglia1∗, Jessica B. Hamrick1 , Victor Bapst1 ,

Alvaro Sanchez-Gonzalez1 , Vinicius Zambaldi1 , Mateusz Malinowski1 ,
Andrea Tacchetti1 , David Raposo1 , Adam Santoro1 , Ryan Faulkner1 ,
Caglar Gulcehre1 , Francis Song1 , Andrew Ballard1 , Justin Gilmer2 ,
George Dahl2 , Ashish Vaswani2 , Kelsey Allen3 , Charles Nash4 ,
arXiv:1806.01261v3 [cs.LG] 17 Oct 2018
Victoria Langston1 , Chris Dyer1 , Nicolas Heess1 ,

Daan Wierstra1 , Pushmeet Kohli1 , Matt Botvinick1 ,
Oriol Vinyals1 , Yujia Li1 , Razvan Pascanu1
1 DeepMind; 2 Google Brain; 3 MIT; 4 University of Edinburgh
Abstract
Artificial intelligence (AI) has undergone a renaissance recently, making major progress in
key domains such as vision, language, control, and decision-making. This has been due, in
part, to cheap data and cheap compute resources, which have fit the natural strengths of deep
learning. However, many defining characteristics of human intelligence, which developed under
much different pressures, remain out of reach for current approaches. In particular, generalizing
beyond one’s experiences—a hallmark of human intelligence from infancy—remains a formidable
challenge for modern AI.
The following is part position paper, part review, and part unification. We argue that
combinatorial generalization must be a top priority for AI to achieve human-like abilities, and that
structured representations and computations are key to realizing this objective. Just as biology
uses nature and nurture cooperatively, we reject the false choice between “hand-engineering”
and “end-to-end” learning, and instead advocate for an approach which benefits from their
complementary strengths. We explore how using relational inductive biases within deep learning
architectures can facilitate learning about entities, relations, and rules for composing them. We
present a new building block for the AI toolkit with a strong relational inductive bias—the graph
network —which generalizes and extends various approaches for neural networks that operate
on graphs, and provides a straightforward interface for manipulating structured knowledge and
producing structured behaviors. We discuss how graph networks can support relational reasoning
and combinatorial generalization, laying the foundation for more sophisticated, interpretable,
and flexible patterns of reasoning. As a companion to this paper, we have also released an
open-source software library for building graph networks, with demonstrations of how to use
them in practice.
1 Introduction
A key signature of human intelligence is the ability to make “infinite use of finite means” (Humboldt,
1836; Chomsky, 1965), in which a small set of elements (such as words) can be productively
composed in limitless ways (such as into new sentences). This reflects the principle of combinatorial
generalization, that is, constructing new inferences, predictions, and behaviors from known building
blocks. Here we explore how to improve modern AI’s capacity for combinatorial generalization by
∗
Corresponding author: peterbattaglia@google.com
1
biasing learning towards structured representations and computations, and in particular, systems
that operate on graphs.
Humans’ capacity for combinatorial generalization depends critically on our cognitive mecha-
nisms for representing structure and reasoning about relations. We represent complex systems as
compositions of entities and their interactions1 (Navon, 1977; McClelland and Rumelhart, 1981;
Plaut et al., 1996; Marcus, 2001; Goodwin and Johnson-Laird, 2005; Kemp and Tenenbaum, 2008),
such as judging whether a haphazard stack of objects is stable (Battaglia et al., 2013). We use
hierarchies to abstract away from fine-grained differences, and capture more general commonalities
between representations and behaviors (Botvinick, 2008; Tenenbaum et al., 2011), such as parts of
an object, objects in a scene, neighborhoods in a town, and towns in a country. We solve novel
problems by composing familiar skills and routines (Anderson, 1982), for example traveling to a
new location by composing familiar procedures and objectives, such as “travel by airplane”, “to
San Diego”, “eat at”, and “an Indian restaurant”. We draw analogies by aligning the relational
structure between two domains and drawing inferences about one based on corresponding knowledge
about the other (Gentner and Markman, 1997; Hummel and Holyoak, 2003).
Kenneth Craik’s “The Nature of Explanation” (1943), connects the compositional structure of
the world to how our internal mental models are organized:
...[a human mental model] has a similar relation-structure to that of the process it
imitates. By ‘relation-structure’ I do not mean some obscure non-physical entity which
attends the model, but the fact that it is a working physical model which works in the
same way as the process it parallels... physical reality is built up, apparently, from a few
fundamental types of units whose properties determine many of the properties of the
most complicated phenomena, and this seems to afford a sufficient explanation of the
emergence of analogies between mechanisms and similarities of relation-structure among
these combinations without the necessity of any theory of objective universals. (Craik,
1943, page 51-55)
That is, the world is compositional, or at least, we understand it in compositional terms. When
learning, we either fit new knowledge into our existing structured representations, or adjust the
structure itself to better accommodate (and make use of) the new and the old (Tenenbaum et al.,
2006; Griffiths et al., 2010; Ullman et al., 2017).
The question of how to build artificial systems which exhibit combinatorial generalization has
been at the heart of AI since its origins, and was central to many structured approaches, including
logic, grammars, classic planning, graphical models, causal reasoning, Bayesian nonparametrics, and
probabilistic programming (Chomsky, 1957; Nilsson and Fikes, 1970; Pearl, 1986, 2009; Russell and
Norvig, 2009; Hjort et al., 2010; Goodman et al., 2012; Ghahramani, 2015). Entire sub-fields have
focused on explicit entity- and relation-centric learning, such as relational reinforcement learning
(Džeroski et al., 2001) and statistical relational learning (Getoor and Taskar, 2007). A key reason
why structured approaches were so vital to machine learning in previous eras was, in part, because
data and computing resources were expensive, and the improved sample complexity afforded by
structured approaches’ strong inductive biases was very valuable.
In contrast with past approaches in AI, modern deep learning methods (LeCun et al., 2015;
Schmidhuber, 2015; Goodfellow et al., 2016) often follow an “end-to-end” design philosophy which
emphasizes minimal a priori representational and computational assumptions, and seeks to avoid
explicit structure and “hand-engineering”. This emphasis has fit well with—and has perhaps been
affirmed by—the current abundance of cheap data and cheap computing resources, which make
1
Whether this entails a “language of thought” (Fodor, 1975) is beyond the scope of this work.
2
trading off sample efficiency for more flexible learning a rational choice. The remarkable and rapid
advances across many challenging domains, from image classification (Krizhevsky et al., 2012;
Szegedy et al., 2017), to natural language processing (Sutskever et al., 2014; Bahdanau et al., 2015),
to game play (Mnih et al., 2015; Silver et al., 2016; Moravčı́k et al., 2017), are a testament to this
minimalist principle. A prominent example is from language translation, where sequence-to-sequence
approaches (Sutskever et al., 2014; Bahdanau et al., 2015) have proven very effective without using
explicit parse trees or complex relationships between linguistic entities.
Despite deep learning’s successes, however, important critiques (Marcus, 2001; Shalev-Shwartz
et al., 2017; Lake et al., 2017; Lake and Baroni, 2018; Marcus, 2018a,b; Pearl, 2018; Yuille and
Liu, 2018) have highlighted key challenges it faces in complex language and scene understanding,
reasoning about structured data, transferring learning beyond the training conditions, and learning
from small amounts of experience. These challenges demand combinatorial generalization, and so it
is perhaps not surprising that an approach which eschews compositionality and explicit structure
struggles to meet them.
When deep learning’s connectionist (Rumelhart et al., 1987) forebears were faced with analogous
critiques from structured, symbolic positions (Fodor and Pylyshyn, 1988; Pinker and Prince, 1988),
there was a constructive effort (Bobrow and Hinton, 1990; Marcus, 2001) to address the challenges
directly and carefully. A variety of innovative sub-symbolic approaches for representing and reasoning
about structured objects were developed in domains such as analogy-making, linguistic analysis,
symbol manipulation, and other forms of relational reasoning (Smolensky, 1990; Hinton, 1990;
Pollack, 1990; Elman, 1991; Plate, 1995; Eliasmith, 2013), as well as more integrative theories for
how the mind works (Marcus, 2001). Such work also helped cultivate more recent deep learning
advances which use distributed, vector representations to capture rich semantic content in text
(Mikolov et al., 2013; Pennington et al., 2014), graphs (Narayanan et al., 2016, 2017), algebraic and
logical expressions (Allamanis et al., 2017; Evans et al., 2018), and programs (Devlin et al., 2017;
Chen et al., 2018b).
We suggest that a key path forward for modern AI is to commit to combinatorial generalization
as a top priority, and we advocate for integrative approaches to realize this goal. Just as biology does
not choose between nature versus nurture—it uses nature and nurture jointly, to build wholes which
are greater than the sums of their parts—we, too, reject the notion that structure and flexibility are
somehow at odds or incompatible, and embrace both with the aim of reaping their complementary
strengths. In the spirit of numerous recent examples of principled hybrids of structure-based methods
and deep learning (e.g., Reed and De Freitas, 2016; Garnelo et al., 2016; Ritchie et al., 2016; Wu
et al., 2017; Denil et al., 2017; Hudson and Manning, 2018), we see great promise in synthesizing
new techniques by drawing on the full AI toolkit and marrying the best approaches from today
with those which were essential during times when data and computation were at a premium.
Recently, a class of models has arisen at the intersection of deep learning and structured
approaches, which focuses on approaches for reasoning about explicitly structured data, in particular
graphs (e.g. Scarselli et al., 2009b; Bronstein et al., 2017; Gilmer et al., 2017; Wang et al., 2018c; Li
et al., 2018; Kipf et al., 2018; Gulcehre et al., 2018). What these approaches all have in common
is a capacity for performing computation over discrete entities and the relations between them.
What sets them apart from classical approaches is how the representations and structure of the
entities and relations—and the corresponding computations—can be learned, relieving the burden
of needing to specify them in advance. Crucially, these methods carry strong relational inductive
biases, in the form of specific architectural assumptions, which guide these approaches towards
learning about entities and relations (Mitchell, 1980), which we, joining many others (Spelke et al.,
1992; Spelke and Kinzler, 2007; Marcus, 2001; Tenenbaum et al., 2011; Lake et al., 2017; Lake and
Baroni, 2018; Marcus, 2018b), suggest are an essential ingredient for human-like intelligence.
3
Box 1: Relational reasoning
We define structure as the product of composing a set of known building blocks. “Structured
representations” capture this composition (i.e., the arrangement of the elements) and “structured
computations” operate over the elements and their composition as a whole. Relational reasoning,
then, involves manipulating structured representations of entities and relations, using rules
for how they can be composed. We use these terms to capture notions from cognitive science,
theoretical computer science, and AI, as follows:
◦ An entity is an element with attributes, such as a physical object with a size and mass.
◦ A relation is a property between entities. Relations between two objects might include
same size as, heavier than, and distance from. Relations can have attributes as
well. The relation more than X times heavier than takes an attribute, X, which
determines the relative weight threshold for the relation to be true vs. false. Relations
can also be sensitive to the global context. For a stone and a feather, the relation falls
with greater acceleration than depends on whether the context is in air vs. in a
vacuum. Here we focus on pairwise relations between entities.
◦ A rule is a function (like a non-binary logical predicate) that maps entities and relations
to other entities and relations, such as a scale comparison like is entity X large? and
is entity X heavier than entity Y?. Here we consider rules which take one or two
arguments (unary and binary), and return a unary property value.
As an illustrative example of relational reasoning in machine learning, graphical models (Pearl,

1988; Koller and Friedman, 2009) can represent complex joint distributions by making explicit
random conditional independences among random variables. Such models have been very
successful because they capture the sparse structure which underlies many real-world generative
processes and because they support efficient algorithms for learning and reasoning. For example,
hidden Markov models constrain latent states to be conditionally independent of others given
the state at the previous time step, and observations to be conditionally independent given the
latent state at the current time step, which are well-matched to the relational structure of many
real-world causal processes. Explicitly expressing the sparse dependencies among variables
provides for various efficient inference and reasoning algorithms, such as message-passing, which
apply a common information propagation procedure across localities within a graphical model,
resulting in a composable, and partially parallelizable, reasoning procedure which can be applied
to graphical models of different sizes and shape.
In the remainder of the paper, we examine various deep learning methods through the lens of
their relational inductive biases, showing that existing methods often carry relational assumptions
which are not always explicit or immediately evident. We then present a general framework for
entity- and relation-based reasoning—which we term graph networks—for unifying and extending
existing methods which operate on graphs, and describe key design principles for building powerful
architectures using graph networks as building blocks. We have also released an open-source library
for building graph networks, which can be found here: github.com/deepmind/graph nets.
2 Relational inductive biases

Many approaches in machine learning and AI which have a capacity for relational reasoning
4
Box 2: Inductive biases
Learning is the process of apprehending useful knowledge by observing and interacting with the
world. It involves searching a space of solutions for one expected to provide a better explanation
of the data or to achieve higher rewards. But in many cases, there are multiple solutions which
are equally good (Goodman, 1955). An inductive bias allows a learning algorithm to prioritize
one solution (or interpretation) over another, independent of the observed data (Mitchell,
1980). In a Bayesian model, inductive biases are typically expressed through the choice and
parameterization of the prior distribution (Griffiths et al., 2010). In other contexts, an inductive
bias might be a regularization term (McClelland, 1994) added to avoid overfitting, or it might
be encoded in the architecture of the algorithm itself. Inductive biases often trade flexibility
for improved sample complexity and can be understood in terms of the bias-variance tradeoff
(Geman et al., 1992). Ideally, inductive biases both improve the search for solutions without
substantially diminishing performance, as well as help find solutions which generalize in a
desirable way; however, mismatched inductive biases can also lead to suboptimal performance
by introducing constraints that are too strong.
Inductive biases can express assumptions about either the data-generating process or the space
of solutions. For example, when fitting a 1D function to data, linear least squares follows
the constraint that the approximating function be a linear model, and approximation errors
should be minimal under a quadratic penalty. This reflects an assumption that the data-
generating process can be explained simply, as a line process corrupted by additive Gaussian
noise. Similarly, L2 regularization prioritizes solutions whose parameters have small values,
and can induce unique solutions and global structure to otherwise ill-posed problems. This can
be interpreted as an assumption about the learning process: that searching for good solutions
is easier when there is less ambiguity among solutions. Note, these assumptions need not be
explicit—they reflect interpretations of how a model or algorithm interfaces with the world.
(Box 1) use a relational inductive bias. While not a precise, formal definition, we use this term to
refer generally to inductive biases (Box 2) which impose constraints on relationships and interactions
among entities in a learning process.
Creative new machine learning architectures have rapidly proliferated in recent years, with
(perhaps not surprisingly given the thesis of this paper) practitioners often following a design pattern
of composing elementary building blocks to form more complex, deep2 computational hierarchies and
graphs3 . Building blocks such as “fully connected” layers are stacked into “multilayer perceptrons”
(MLPs), “convolutional layers” are stacked into “convolutional neural networks” (CNNs), and a
standard recipe for an image processing network is, generally, some variety of CNN composed with
a MLP. This composition of layers provides a particular type of relational inductive bias—that
of hierarchical processing—in which computations are performed in stages, typically resulting in
increasingly long range interactions among information in the input signal. As we explore below, the
building blocks themselves also carry various relational inductive biases (Table 1). Though beyond
the scope of this paper, various non-relational inductive biases are used in deep learning as well: for
example, activation non-linearities, weight decay, dropout (Srivastava et al., 2014), batch and layer
normalization (Ioffe and Szegedy, 2015; Ba et al., 2016), data augmentation, training curricula, and
optimization algorithms all impose constraints on the trajectory and outcome of learning.
2
This pattern of composition in depth is ubiquitous in deep learning, and is where the “deep” comes from.
3
Recent methods (Liu et al., 2018) even automate architecture construction via learned graph editing procedures.
5
Component Entities Relations Rel. inductive bias Invariance
Fully connected Units All-to-all Weak -
Convolutional Grid elements Local Locality Spatial translation
Recurrent Timesteps Sequential Sequentiality Time translation
Graph network Nodes Edges Arbitrary Node, edge permutations
Table 1: Various relational inductive biases in standard deep learning components. See also Section 2.
To explore the relational inductive biases expressed within various deep learning methods, we
must identify several key ingredients, analogous to those in Box 1: what are the entities, what are
the relations, and what are the rules for composing entities and relations, and computing their
implications? In deep learning, the entities and relations are typically expressed as distributed
representations, and the rules as neural network function approximators; however, the precise forms
of the entities, relations, and rules vary between architectures. To understand these differences
between architectures, we can further ask how each supports relational reasoning by probing:
◦ The arguments to the rule functions (e.g., which entities and relations are provided as input).
◦ How the rule function is reused, or shared, across the computational graph (e.g., across different
entities and relations, across different time or processing steps, etc.).
◦ How the architecture defines interactions versus isolation among representations (e.g., by
applying rules to draw conclusions about related entities, versus processing them separately).
2.1 Relational inductive biases in standard deep learning building blocks

2.1.1 Fully connected layers
Perhaps the most common building block is a fully connected layer (Rosenblatt, 1961). Typically
implemented as a non-linear vector-valued function of vector inputs, each element, or “unit”, of
the output vector is the dot product between a weight vector, followed by an added bias term, and
finally a non-linearity such as a rectified linear unit (ReLU). As such, the entities are the units in
the network, the relations are all-to-all (all units in layer i are connected to all units in layer j),
and the rules are specified by the weights and biases. The argument to the rule is the full input
signal, there is no reuse, and there is no isolation of information (Figure 1a). The implicit relational
inductive bias in a fully connected layer is thus very weak: all input units can interact to determine
any output unit’s value, independently across outputs (Table 1).
2.1.2 Convolutional layers
Another common building block is a convolutional layer (Fukushima, 1980; LeCun et al., 1989). It is
implemented by convolving an input vector or tensor with a kernel of the same rank, adding a bias
term, and applying a point-wise non-linearity. The entities here are still individual units (or grid
elements, e.g. pixels), but the relations are sparser. The differences between a fully connected layer
and a convolutional layer impose some important relational inductive biases: locality and translation
invariance (Figure 1b). Locality reflects that the arguments to the relational rule are those entities in
close proximity with one another in the input signal’s coordinate space, isolated from distal entities.
Translation invariance reflects reuse of the same rule across localities in the input. These biases
are very effective for processing natural image data because there is high covariance within local
6
Sharing in time
Sharing in space
(a) Fully connected (b) Convolutional (c) Recurrent
Figure 1: Reuse and sharing in common deep learning building blocks. (a) Fully connected layer,
in which all weights are independent, and there is no sharing. (b) Convolutional layer, in which
a local kernel function is reused multiple times across the input. Shared weights are indicated by
arrows with the same color. (c) Recurrent layer, in which the same function is reused across different
processing steps.
neighborhoods, which diminishes with distance, and because the statistics are mostly stationary
across an image (Table 1).
2.1.3 Recurrent layers

A third common building block is a recurrent layer (Elman, 1990), which is implemented over a
sequence of steps. Here, we can view the inputs and hidden states at each processing step as the
entities, and the Markov dependence of one step’s hidden state on the previous hidden state and
the current input, as the relations. The rule for combining the entities takes a step’s inputs and
hidden state as arguments to update the hidden state. The rule is reused over each step (Figure 1c),
which reflects the relational inductive bias of temporal invariance (similar to a CNN’s translational
invariance in space). For example, the outcome of some physical sequence of events should not
depend on the time of day. RNNs also carry a bias for locality in the sequence via their Markovian
structure (Table 1).
2.2 Computations over sets and graphs

While the standard deep learning toolkit contains methods with various forms of relational inductive
biases, there is no “default” deep learning component which operates on arbitrary relational structure.
We need models with explicit representations of entities and relations, and learning algorithms which
find rules for computing their interactions, as well as ways of grounding them in data. Importantly,
entities in the world (such as objects and agents) do not have a natural order; rather, orderings
can be defined by the properties of their relations. For example, the relations between the sizes of
a set of objects can potentially be used to order them, as can their masses, ages, toxicities, and
prices. Invariance to ordering—except in the face of relations—is a property that should ideally be
reflected by a deep learning component for relational reasoning.
Sets are a natural representation for systems which are described by entities whose order is
undefined or irrelevant; in particular, their relational inductive bias does not come from the presence
of something, but rather from the absence. For illustration, consider the task of predicting the center
7
(a) Molecule (b) Mass-Spring System
H H
(c) n-body System (d) Rigid Body System
(e) Sentence and Parse Tree (f) Image and Fully-Connected Scene Graph
The brown
dog jumped. The jumped
brown dog
Figure 2: Different graph representations. (a) A molecule, in which each atom is represented as a
node and edges correspond to bonds (e.g. Duvenaud et al., 2015). (b) A mass-spring system, in
which the rope is defined by a sequence of masses which are represented as nodes in the graph (e.g.
Battaglia et al., 2016; Chang et al., 2017). (c) A n-body system, in which the bodies are nodes and
the underlying graph is fully connected (e.g. Battaglia et al., 2016; Chang et al., 2017). (d) A rigid
body system, in which the balls and walls are nodes, and the underlying graph defines interactions
between the balls and between the balls and the walls (e.g. Battaglia et al., 2016; Chang et al., 2017).
(e) A sentence, in which the words correspond to leaves in a tree, and the other nodes and edges
could be provided by a parser (e.g. Socher et al., 2013). Alternately, a fully connected graph could
be used (e.g. Vaswani et al., 2017). (f) An image, which can be decomposed into image patches
corresponding to nodes in a fully connected graph (e.g. Santoro et al., 2017; Wang et al., 2018c).
of mass of a solar system comprised of n planets, whose attributes (e.g., mass, position, velocity,
etc.) are denoted by {x1 , x2 , . . . , xn }. For such a computation, the order in which we consider the
planets does not matter because the state can be described solely in terms of aggregated, averaged
quantities. However, if we were to use a MLP for this task, having learned the prediction for a
particular input (x1 , x2 , . . . , xn ) would not necessarily transfer to making a prediction for the same
inputs under a different ordering (xn , x1 , . . . , x2 ). Since there are n! such possible permutations, in
the worst case, the MLP could consider each ordering as fundamentally different, and thus require
an exponential number of input/output training examples to learn an approximating function.
A natural way to handle such combinatorial explosion is to only allow the prediction to depend
on symmetric functions of the inputs’ attributes. This might mean computing shared per-object
8
features {f (x1 ), . . . , f (xn )} which are then aggregated in a symmetric way (for example, by taking
their mean). Such an approach is the essence of the Deep Sets and related models (Zaheer et al.,
2017; Edwards and Storkey, 2016; Pevnỳ and Somol, 2017), which we explore further in Section 4.2.3.
Of course, permutation invariance is not the only important form of underlying structure in
many problems. For example, each object in a set may be affected by pairwise interactions with
the other objects in the set (Hartford et al., 2018). In our planets scenario, consider now the
task of predicting each individual planet’s position after a time interval, ∆t. In this case, using
aggregated, averaged information is not enough because the movement of each planet depends on
the forces the P other planets are exerting on it. Instead, we could compute the state of each object
0
as xi = f (xi , j g(xi , xj )), where g could compute the force induced by the j-th planet on the
i-th planet, and f could compute the future state of the i-th planet which results from the forces
and dynamics. The fact that we use the same g everywhere is again a consequence of the global
permutation invariance of the system; however, it also supports a different relational structure
because g now takes two arguments rather than one.4
The above solar system examples illustrate two relational structures: one in which there are
no relations, and one which consists of all pairwise relations. Many real-world systems (such as
in Figure 2) have a relational structure somewhere in between these two extremes, however, with
some pairs of entities possessing a relation and others lacking one. In our solar system example, if
the system instead consists of the planets and their moons, one may be tempted to approximate
it by neglecting the interactions between moons of different planets. In P practice, this means
computing interactions only between some pairs of objects, i.e. x0i = f (xi , j∈δ(i) g(xi , xj )), where
δ(i) ⊆ {1, . . . , n} is a neighborhood around node i. This corresponds to a graph, in that the i-th
object only interacts with a subset of the other objects, described by its neighborhood. Note, the
updated states still do not depend in the order in which we describe the neighborhood.5
Graphs, generally, are a representation which supports arbitrary (pairwise) relational struc-
ture, and computations over graphs afford a strong relational inductive bias beyond that which
convolutional and recurrent layers can provide.
3 Graph networks
Neural networks that operate on graphs, and structure their computations accordingly, have been
developed and explored extensively for more than a decade under the umbrella of “graph neural
networks” (Gori et al., 2005; Scarselli et al., 2005, 2009a; Li et al., 2016), but have grown rapidly
in scope and popularity in recent years. We survey the literature on these methods in the next
sub-section (3.1). Then in the remaining sub-sections, we present our graph networks framework,
which generalizes and extends several lines of work in this area.
3.1 Background
Models in the graph neural network family (Gori et al., 2005; Scarselli et al., 2005, 2009a; Li et al.,
2016) have been explored in a diverse range of problem domains, across supervised, semi-supervised,
unsupervised, and reinforcement learning settings. They have been effective at tasks thought to
have rich relational structure, such as visual scene understanding tasks (Raposo et al., 2017; Santoro
4
We could extend this same analysis to increasingly entangled structures that depend on relations among triplets
(i.e., g(xi , xj , xk )), quartets, and so on. We note that if we restrict these functions to only operate on subsets of xi
which are spatially close, then we end back up with something resembling CNNs. In the most entangled sense, where
there is a single relation function g(x1 , . . . , xn ), we end back up with a construction similar to a fully connected layer.
5
The invariance which this model enforces is the invariance under isomorphism of the graph.
9
et al., 2017) and few-shot learning (Garcia and Bruna, 2018). They have also been used to learn
the dynamics of physical systems (Battaglia et al., 2016; Chang et al., 2017; Watters et al., 2017;
van Steenkiste et al., 2018; Sanchez-Gonzalez et al., 2018) and multi-agent systems (Sukhbaatar
et al., 2016; Hoshen, 2017; Kipf et al., 2018), to reason about knowledge graphs (Bordes et al., 2013;
Oñoro-Rubio et al., 2017; Hamaguchi et al., 2017), to predict the chemical properties of molecules
(Duvenaud et al., 2015; Gilmer et al., 2017), to predict traffic on roads (Li et al., 2017; Cui et al.,
2018), to classify and segment images and videos (Wang et al., 2018c; Hu et al., 2017) and 3D
meshes and point clouds (Wang et al., 2018d), to classify regions in images (Chen et al., 2018a), to
perform semi-supervised text classification (Kipf and Welling, 2017), and in machine translation
(Vaswani et al., 2017; Shaw et al., 2018; Gulcehre et al., 2018). They have been used within both
model-free (Wang et al., 2018b) and model-based (Hamrick et al., 2017; Pascanu et al., 2017;
Sanchez-Gonzalez et al., 2018) continuous control, for model-free reinforcement learning (Hamrick
et al., 2018; Zambaldi et al., 2018), and for more classical approaches to planning (Toyer et al.,
2017).
Many traditional computer science problems, which involve reasoning about discrete entities and
structure, have also been explored with graph neural networks, such as combinatorial optimization
(Bello et al., 2016; Nowak et al., 2017; Dai et al., 2017), boolean satisfiability (Selsam et al., 2018),
program representation and verification (Allamanis et al., 2018; Li et al., 2016), modeling cellular
automata and Turing machines (Johnson, 2017), and performing inference in graphical models
(Yoon et al., 2018). Recent work has also focused on building generative models of graphs (Li et al.,
2018; De Cao and Kipf, 2018; You et al., 2018; Bojchevski et al., 2018), and unsupervised learning of
graph embeddings (Perozzi et al., 2014; Tang et al., 2015; Grover and Leskovec, 2016; Garcı́a-Durán
and Niepert, 2017).
The works cited above are by no means an exhaustive list, but provide a representative cross-
section of the breadth of domains for which graph neural networks have proven useful. We point
interested readers to a number of existing reviews which examine the body of work on graph neural
networks in more depth. In particular, Scarselli et al. (2009a) provides an authoritative overview of
early graph neural network approaches. Bronstein et al. (2017) provides an excellent survey of deep
learning on non-Euclidean data, and explores graph neural nets, graph convolution networks, and
related spectral approaches. Recently, Gilmer et al. (2017) introduced the message-passing neural
network (MPNN), which unified various graph neural network and graph convolutional network
approaches (Monti et al., 2017; Bruna et al., 2014; Henaff et al., 2015; Defferrard et al., 2016;
Niepert et al., 2016; Kipf and Welling, 2017; Bronstein et al., 2017) by analogy to message-passing
in graphical models. In a similar vein, Wang et al. (2018c) introduced the non-local neural network
(NLNN), which unified various “self-attention”-style methods (Vaswani et al., 2017; Hoshen, 2017;
Veličković et al., 2018) by analogy to methods from computer vision and graphical models for
capturing long range dependencies in signals.
3.2 Graph network (GN) block

We now present our graph networks (GN) framework, which defines a class of functions for
relational reasoning over graph-structured representations. Our GN framework generalizes and
extends various graph neural network, MPNN, and NLNN approaches (Scarselli et al., 2009a; Gilmer
et al., 2017; Wang et al., 2018c), and supports constructing complex architectures from simple
building blocks. Note, we avoided using the term “neural” in the “graph network” label to reflect
that they can be implemented with functions other than neural networks, though here our focus is
on neural network implementations.
The main unit of computation in the GN framework is the GN block, a “graph-to-graph” module
10
Box 3: Our definition of “graph”
Attributes
vi <latexit sha1_base64="4ton1cC0/WpHTbJYOP5RCFkc+ww=">AAAB83icbVDLSsNAFL3xWeur6tLNYBG6KokIuiy4cVnBPqAJZTK9aYdOJmFmIpTQ33DjQhG3/ow7/8ZJm4W2Hhg4nHMv98wJU8G1cd1vZ2Nza3tnt7JX3T84PDqunZx2dZIphh2WiET1Q6pRcIkdw43AfqqQxqHAXji9K/zeEyrNE/loZikGMR1LHnFGjZV8P6ZmEkY5zofTYa3uNt0FyDrxSlKHEu1h7csfJSyLURomqNYDz01NkFNlOBM4r/qZxpSyKR3jwFJJY9RBvsg8J5dWGZEoUfZJQxbq742cxlrP4tBOFhn1qleI/3mDzES3Qc5lmhmUbHkoygQxCSkKICOukBkxs4QyxW1WwiZUUWZsTVVbgrf65XXSvWp6btN7uK63GmUdFTiHC2iABzfQgntoQwcYpPAMr/DmZM6L8+58LEc3nHLnDP7A+fwBXXGRzQ==</latexit>
ek u u
vi
ek
v sk v rk
Here we use “graph” to mean a directed, attributed multi-graph with a global attribute. In our
terminology, a node is denoted as vi , an edge as ek , and the global attributes as u. We also use
sk and rk to indicate the indices of the sender and receiver nodes (see below), respectively, for
edge k. To be more precise, we define these terms as:
Directed : one-way edges, from a “sender” node to a “receiver” node.

Attribute : properties that can be encoded as a vector, set, or even another graph.
Attributed : edges and vertices have attributes associated with them.
Global attribute : a graph-level attribute.
Multi-graph : there can be more than one edge between vertices, including self-edges.
Figure 2 shows a variety of different types of graphs corresponding to real data that we may be
interested in modeling, including physical systems, molecules, images, and text.
which takes a graph as input, performs computations over the structure, and returns a graph as
output. As described in Box 3, entities are represented by the graph’s nodes, relations by the edges,
and system-level properties by global attributes. The GN framework’s block organization emphasizes
customizability and synthesizing new architectures which express desired relational inductive biases.
The key design principles are: Flexible representations (see Section 4.1); Configurable within-block
structure (see Section 4.2); and Composable multi-block architectures (see Section 4.3).
We introduce a motivating example to help make the GN formalism more concrete. Consider
predicting the movements a set of rubber balls in an arbitrary gravitational field, which, instead of
bouncing against one another, each have one or more springs which connect them to some (or all) of
the others. We will refer to this running example throughout the definitions below, to motivate the
graph representation and the computations operating over it. Figure 2 depicts some other common
scenarios that can be represented by graphs and reasoned over using graph networks.
3.2.1 Definition of “graph”
Within our GN framework, a graph is defined as a 3-tuple G = (u, V, E) (see Box 3 for details of
graph representations). The u is a global attribute; for example, u might represent the gravitational
field. The V = {vi }i=1:N v is the set of nodes (of cardinality N v ), where each vi is a node’s attribute.
For example, V might represent each ball, with attributes for position, velocity, and mass. The
E = {(ek , rk , sk )}k=1:N e is the set of edges (of cardinality N e ), where each ek is the edge’s attribute,
rk is the index of the receiver node, and sk is the index of the sender node. For example, E might
represent the presence of springs between different balls, and their corresponding spring constants.
11
Algorithm 1 Steps of computation in a full GN block.
function GraphNetwork(E, V , u)
for k ∈ {1 . . . N e } do
e0k ← φe (ek , vrk , vsk , u) . 1. Compute updated edge attributes
end for
for i ∈ {1 . . . N n } do
let Ei0 = {(e0k , rk , sk )}rk =i, k=1:N e
ē0i ← ρe→v (Ei0 ) . 2. Aggregate edge attributes per node
0 v 0
vi ← φ (ēi , vi , u) . 3. Compute updated node attributes
end for
let V 0 = {v0 }i=1:N v
let E 0 = {(e0k , rk , sk )}k=1:N e
ē0 ← ρe→u (E 0 ) . 4. Aggregate edge attributes globally
0
v̄ ← ρ v→u (V )0 . 5. Aggregate node attributes globally
u0 ← φu (ē0 , v̄0 , u) . 6. Compute updated global attribute
return (E 0 , V 0 , u0 )
end function
3.2.2 Internal structure of a GN block
A GN block contains three “update” functions, φ, and three “aggregation” functions, ρ,
e0k = φe (ek , vrk , vsk , u) ē0i = ρe→v Ei0

vi0 = φv ē0i , vi , u ē0 = ρe→u E 0

(1)
u0 = φu ē0 , v̄0 , u 0 v→u 0

v̄ = ρ V
where Ei0 = {(e0k , rk , sk )}rk =i, k=1:N e , V 0 = {vi0 }i=1:N v , and E 0 = i Ei0 = {(e0k , rk , sk )}k=1:N e .
S
The φe is mapped across all edges to compute per-edge updates, the φv is mapped across all
nodes to compute per-node updates, and the φu is applied once as the global update. The ρ
functions each take a set as input, and reduce it to a single element which represents the aggregated
information. Crucially, the ρ functions must be invariant to permutations of their inputs, and should
take variable numbers of arguments (e.g., elementwise summation, mean, maximum, etc.).
3.2.3 Computational steps within a GN block
When a graph, G, is provided as input to a GN block, the computations proceed from the edge, to
the node, to the global level. Figure 3 shows a depiction of which graph elements are involved in
each of these computations, and Figure 4a shows a full GN block, with its update and aggregation
functions. Algorithm 1 shows the following steps of computation:
1. φe is applied per edge, with arguments (ek , vrk , vsk , u), and returns e0k . In our springs example,
this might correspond to the forces or potential energies between two connected balls. The
set of resulting per-edge outputs for each node, i, is, Ei0 = {(e0k , rk , sk )}rk =i, k=1:N e . And
E 0 = i Ei0 = {(e0k , rk , sk )}k=1:N e is the set of all per-edge outputs.
S
2. ρe→v is applied to Ei0 , and aggregates the edge updates for edges that project to vertex i, into
ē0i , which will be used in the next step’s node update. In our running example, this might
correspond to summing all the forces or potential energies acting on the ith ball.
12
vi <latexit sha1_base64="a1hco1MShws4KpmpFnenOcfEqyc=">AAAB9HicdVDLSgMxFM34rPVVdekmWMSuhkxtsd0V3LisYB/QDiWT3mlDMw+TTKEM/Q43LhRx68e482/MtBVU9EDgcM693JPjxYIrTciHtba+sbm1ndvJ7+7tHxwWjo7bKkokgxaLRCS7HlUgeAgtzbWAbiyBBp6Ajje5zvzOFKTiUXinZzG4AR2F3OeMaiO5/YDqseenML8YTAaFIrGdCilXy5jY1XqdVOqG1KqXpEywY5MFimiF5qDw3h9GLAkg1ExQpXoOibWbUqk5EzDP9xMFMWUTOoKeoSENQLnpIvQcnxtliP1ImhdqvFC/b6Q0UGoWeGYyC6l+e5n4l9dLtF9zUx7GiYaQLQ/5icA6wlkDeMglMC1mhlAmucmK2ZhKyrTpKW9K+Pop/p+0y7ZjurqtFBulVR05dIrOUAk56Ao10A1qohZi6B49oCf0bE2tR+vFel2OrlmrnRP0A9bbJyRskkI=</latexit>
e0k <latexit sha1_base64="Wl/NKcf+4FQq41kPZqpr8GSpKP8=">AAAB8XicbVDLSsNAFL2pr1pfVZduBovQVUlE0GXBjcsK9oFtKJPpTTt0MgkzE6GE/oUbF4q49W/c+TdO2iy09cDA4Zx7mXNPkAiujet+O6WNza3tnfJuZW//4PCoenzS0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0Nve7T6g0j+WDmSXoR3QsecgZNVZ6HETUTIIwS+fDas1tuAuQdeIVpAYFWsPq12AUszRCaZigWvc9NzF+RpXhTOC8Mkg1JpRN6Rj7lkoaofazReI5ubDKiISxsk8aslB/b2Q00noWBXYyT6hXvVz8z+unJrzxMy6T1KBky4/CVBATk/x8MuIKmREzSyhT3GYlbEIVZcaWVLEleKsnr5POZcNzG979Va1ZL+oowxmcQx08uIYm3EEL2sBAwjO8wpujnRfn3flYjpacYucU/sD5/AHw2JD/</latexit>
<latexit
u vi0 <latexit sha1_base64="TmBm7ikN3ChoJpDcsfwhm1T5rLk=">AAAB9HicbVDLSgMxFL1TX7W+qi7dBIvYVZkRQZcFNy4r2Ae0Q8mkd9rQTGZMMoUy9DvcuFDErR/jzr8x03ahrQcCh3Pu5Z6cIBFcG9f9dgobm1vbO8Xd0t7+weFR+fikpeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPxXe63J6g0j+WjmSboR3QoecgZNVbyexE1oyDMcHbZH/fLFbfmzkHWibckFVii0S9/9QYxSyOUhgmqdddzE+NnVBnOBM5KvVRjQtmYDrFrqaQRaj+bh56RC6sMSBgr+6Qhc/X3RkYjradRYCfzkHrVy8X/vG5qwls/4zJJDUq2OBSmgpiY5A2QAVfIjJhaQpniNithI6ooM7anki3BW/3yOmld1Ty35j1cV+rVZR1FOINzqIIHN1CHe2hAExg8wTO8wpszcV6cd+djMVpwljun8AfO5w/CAJH+</latexit>
e0k <latexit sha1_base64="Wl/NKcf+4FQq41kPZqpr8GSpKP8=">AAAB8XicbVDLSsNAFL2pr1pfVZduBovQVUlE0GXBjcsK9oFtKJPpTTt0MgkzE6GE/oUbF4q49W/c+TdO2iy09cDA4Zx7mXNPkAiujet+O6WNza3tnfJuZW//4PCoenzS0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0Nve7T6g0j+WDmSXoR3QsecgZNVZ6HETUTIIwS+fDas1tuAuQdeIVpAYFWsPq12AUszRCaZigWvc9NzF+RpXhTOC8Mkg1JpRN6Rj7lkoaofazReI5ubDKiISxsk8aslB/b2Q00noWBXYyT6hXvVz8z+unJrzxMy6T1KBky4/CVBATk/x8MuIKmREzSyhT3GYlbEIVZcaWVLEleKsnr5POZcNzG979Va1ZL+oowxmcQx08uIYm3EEL2sBAwjO8wpujnRfn3flYjpacYucU/sD5/AHw2JD/</latexit>
<latexit
u v0i <latexit sha1_base64="Iztn6Umi7rLG5lNF0JpW0x6J+s0=">AAAB9HicbVBNS8NAEN34WetX1aOXxSL2VBIR9Fjw4rGC/YA2lM120i7dbOLupFhCf4cXD4p49cd489+4bXPQ1gcDj/dmmJkXJFIYdN1vZ219Y3Nru7BT3N3bPzgsHR03TZxqDg0ey1i3A2ZACgUNFCihnWhgUSChFYxuZ35rDNqIWD3gJAE/YgMlQsEZWsnvIjxhEGYwveiNeqWyW3XnoKvEy0mZ5Kj3Sl/dfszTCBRyyYzpeG6CfsY0Ci5hWuymBhLGR2wAHUsVi8D42fzoKT23Sp+GsbalkM7V3xMZi4yZRIHtjBgOzbI3E//zOimGN34mVJIiKL5YFKaSYkxnCdC+0MBRTixhXAt7K+VDphlHm1PRhuAtv7xKmpdVz61691flWiWPo0BOyRmpEI9ckxq5I3XSIJw8kmfySt6csfPivDsfi9Y1J585IX/gfP4A6+WSGQ==</latexit>
sha1_base64="hP+6LrUf2d3tZaldqaQQvEKMXyw=">AAAB2XicbZDNSgMxFIXv1L86Vq1rN8EiuCozbnQpuHFZwbZCO5RM5k4bmskMyR2hDH0BF25EfC93vo3pz0JbDwQ+zknIvSculLQUBN9ebWd3b/+gfugfNfzjk9Nmo2fz0gjsilzl5jnmFpXU2CVJCp8LgzyLFfbj6f0i77+gsTLXTzQrMMr4WMtUCk7O6oyaraAdLMW2IVxDC9YaNb+GSS7KDDUJxa0dhEFBUcUNSaFw7g9LiwUXUz7GgUPNM7RRtRxzzi6dk7A0N+5oYkv394uKZ9bOstjdzDhN7Ga2MP/LBiWlt1EldVESarH6KC0Vo5wtdmaJNChIzRxwYaSblYkJN1yQa8Z3HYSbG29D77odBu3wMYA6nMMFXEEIN3AHD9CBLghI4BXevYn35n2suqp569LO4I+8zx84xIo4</latexit>
sha1_base64="ywz7v1q7Yrl4nBX/+QcnkaM0kGo=">AAAB6XicbZBLSwMxFIXv+Kz1Vd26CRbRVZlxo0vBjcsK9gHtUDLpnRqayYzJnWIZ+jvcuFDEP+TOf2P6WGjrgcDHOQn35kSZkpZ8/9tbW9/Y3Nou7ZR39/YPDitHe02b5kZgQ6QqNe2IW1RSY4MkKWxnBnkSKWxFw9tp3hqhsTLVDzTOMEz4QMtYCk7OCruEzxTFBU7Oe8NeperX/JnYKgQLqMJC9V7lq9tPRZ6gJqG4tZ3AzygsuCEpFE7K3dxixsWQD7DjUPMEbVjMlp6wM+f0WZwadzSxmfv7RcETa8dJ5G4mnB7tcjY1/8s6OcXXYSF1lhNqMR8U54pRyqYNsL40KEiNHXBhpNuViUduuCDXU9mVECx/eRWal7XArwX3PpTgBE7hAgK4ghu4gzo0QMATvMAbvHsj79X7mNe15i16O4Y/8j5/AKfckNQ=</latexit>
sha1_base64="Wxt2EGaSkqmyg6rX9KQvpR9rldE=">AAAB9HicbVA9TwJBEN3DL8Qv1NJmIzFSkTsbLUlsLDGRjwQuZG+Zgw17e+fuHJFc+B02Fhpj64+x89+4wBUKvmSSl/dmMjMvSKQw6LrfTmFjc2t7p7hb2ts/ODwqH5+0TJxqDk0ey1h3AmZACgVNFCihk2hgUSChHYxv5357AtqIWD3gNAE/YkMlQsEZWsnvITxhEGYwu+yP++WKW3MXoOvEy0mF5Gj0y1+9QczTCBRyyYzpem6CfsY0Ci5hVuqlBhLGx2wIXUsVi8D42eLoGb2wyoCGsbalkC7U3xMZi4yZRoHtjBiOzKo3F//zuimGN34mVJIiKL5cFKaSYkznCdCB0MBRTi1hXAt7K+UjphlHm1PJhuCtvrxOWlc1z615926lXs3jKJIzck6qxCPXpE7uSIM0CSeP5Jm8kjdn4rw4787HsrXg5DOn5A+czx/qpZIV</latexit>
e0k u0
<latexit sha1_base64="RuJ/WOWmv0qWsx0aAsZGj4qvbr4=">AAAB8nicdVDJSgNBEO2JW4xb1KOXxiDmNMxk0cwt4MVjBLPAZAg9nZ6kSc9Cd40YhnyGFw+KePVrvPk3dhZBRR8UPN6roqqenwiuwLI+jNza+sbmVn67sLO7t39QPDzqqDiVlLVpLGLZ84ligkesDRwE6yWSkdAXrOtPruZ+945JxePoFqYJ80IyinjAKQEtuX1g9+AHWTo7HxRLllm5aFTqFrbMulOrNmqaNCpO1XGwbVoLlNAKrUHxvT+MaRqyCKggSrm2lYCXEQmcCjYr9FPFEkInZMRcTSMSMuVli5Nn+EwrQxzEUlcEeKF+n8hIqNQ09HVnSGCsfntz8S/PTSFoeBmPkhRYRJeLglRgiPH8fzzkklEQU00IlVzfiumYSEJBp1TQIXx9iv8nnYppW6Z9Uys1y6s48ugEnaIystElaqJr1EJtRFGMHtATejbAeDRejNdla85YzRyjHzDePgEJTZGr</latexit>
<latexit sha1_base64="UuhsKP3lpHlY+K0A8uvGImQtNkI=">AAAB83icbVDLSsNAFL2pr1pfVZduBovQVUlE0GXBjcsK9gFNKJPppB06mYSZm0IJ/Q03LhRx68+482+ctllo64GBwzn3cs+cMJXCoOt+O6Wt7Z3dvfJ+5eDw6PikenrWMUmmGW+zRCa6F1LDpVC8jQIl76Wa0ziUvBtO7hd+d8q1EYl6wlnKg5iOlIgEo2gl348pjsMon84HYlCtuQ13CbJJvILUoEBrUP3yhwnLYq6QSWpM33NTDHKqUTDJ5xU/MzylbEJHvG+pojE3Qb7MPCdXVhmSKNH2KSRL9fdGTmNjZnFoJxcZzbq3EP/z+hlGd0EuVJohV2x1KMokwYQsCiBDoTlDObOEMi1sVsLGVFOGtqaKLcFb//Im6Vw3PLfhPd7UmvWijjJcwCXUwYNbaMIDtKANDFJ4hld4czLnxXl3PlajJafYOYc/cD5/AHRgkdw=</latexit>
<latexit sha1_base64="6WjtAQy1eEki3DeLmUkkI9Sk/Os=">AAAB53icbVDLSsNAFL2pr1pfVZduBovQVUlE0GXBjcsK9gFtkMnkph06mYSZiVBCf8CNC0Xc+kvu/BunaRbaemDgcM65zL0nSAXXxnW/ncrG5tb2TnW3trd/cHhUPz7p6SRTDLssEYkaBFSj4BK7hhuBg1QhjQOB/WB6u/D7T6g0T+SDmaXox3QsecQZNVbqPNYbbsstQNaJV5IGlLD5r1GYsCxGaZigWg89NzV+TpXhTOC8Nso0ppRN6RiHlkoao/bzYs85ubBKSKJE2ScNKdTfEzmNtZ7FgU3G1Ez0qrcQ//OGmYlu/JzLNDMo2fKjKBPEJGRxNAm5QmbEzBLKFLe7EjahijJjq6nZErzVk9dJ77LluS3v/qrRbpZ1VOEMzqEJHlxDG+6gA11gEMIzvMKbw50X5935WEYrTjlzCn/gfP4AA6WMYA==</latexit>
sha1_base64="PT7VlVtIO1b4RdkSG9z8jpkhSqk=">AAAB9HicbVBNS8NAEJ34WetX1aOXxSL2VBIR9Fjw4rGC/YA2lM120y7dbOLupFhCf4cXD4p49cd489+4bXPQ1gcDj/dmmJkXJFIYdN1vZ219Y3Nru7BT3N3bPzgsHR03TZxqxhsslrFuB9RwKRRvoEDJ24nmNAokbwWj25nfGnNtRKwecJJwP6IDJULBKFrJ7yJ/wiDMxtOLnuiVym7VnYOsEi8nZchR75W+uv2YpRFXyCQ1puO5CfoZ1SiY5NNiNzU8oWxEB7xjqaIRN342P3pKzq3SJ2GsbSkkc/X3REYjYyZRYDsjikOz7M3E/7xOiuGNnwmVpMgVWywKU0kwJrMESF9ozlBOLKFMC3srYUOqKUObU9GG4C2/vEqal1XPrXr3V+VaJY+jAKdwBhXw4BpqcAd1aACDR3iGV3hzxs6L8+58LFrXnHzmBP7A+fwBAvSSKA==</latexit>
sha1_base64="OfCCjkcIiyDbxmNKxve032U7QH4=">AAAB3HicbVBNS8NAEJ3Ur1qrVq9eFovgqSRe9Ch48VjBfkAbZLOZtEs3m7A7EUroH/DiQRF/lzf/jduPg7Y+GHi8N8PMvChX0pLvf3uVre2d3b3qfu2gfnh03Dipd21WGIEdkanM9CNuUUmNHZKksJ8b5GmksBdN7uZ+7xmNlZl+pGmOYcpHWiZScHJS+6nR9Fv+AmyTBCvShBVc/9cwzkSRoiahuLWDwM8pLLkhKRTOasPCYs7FhI9w4KjmKdqwXNw5YxdOiVmSGVea2EL9PVHy1NppGrnOlNPYrntz8T9vUFByE5ZS5wWhFstFSaEYZWz+NIulQUFq6ggXRrpbmRhzwwW5aGouhGD95U3SvWoFfit48KEKZ3AOlxDANdzCPbShAwJieIE3ePek9+p9LOOqeKvcTuEPvM8f/CqLKA==</latexit>
sha1_base64="PwZ8GjhNs4EFPNOrNCQnWexiUCQ=">AAAB53icbVDLSgMxFL1TX7W+qi7dBIvQVZlxo8uCG5cV7APaQTKZTBuaZIbkjlCG/oAbF4q49Zfc+Tem7Sy0eiBwOOdccu+JMiks+v6XV9nY3Nreqe7W9vYPDo/qxyc9m+aG8S5LZWoGEbVcCs27KFDyQWY4VZHk/Wh6s/D7j9xYkep7nGU8VHSsRSIYRSd1HuoNv+UvQf6SoCQNKOHyn6M4ZbniGpmk1g4DP8OwoAYFk3xeG+WWZ5RN6ZgPHdVUcRsWyz3n5MIpMUlS455GslR/ThRUWTtTkUsqihO77i3E/7xhjsl1WAid5cg1W32U5JJgShZHk1gYzlDOHKHMCLcrYRNqKENXTc2VEKyf/Jf0LluB3wru/Ea7WdZRhTM4hyYEcAVtuIUOdIFBDE/wAq+e8J69N+99Fa145cwp/IL38Q0CZYxc</latexit>
<latexit sha1_base64="eeLXdOBZMDToGpT2JKCAlGanLL8=">AAAB9HicdVDLSgMxFL3js9ZX1aWbYBG7GjK1xXZXcOOygn1AO5RMmmlDMw+TTKEM/Q43LhRx68e482/MtBVU9EDgcM693JPjxYIrjfGHtba+sbm1ndvJ7+7tHxwWjo7bKkokZS0aiUh2PaKY4CFraa4F68aSkcATrONNrjO/M2VS8Si807OYuQEZhdznlGgjuf2A6LHnp9P5xYAPCkVsOxVcrpYRtqv1Oq7UDalVL3EZI8fGCxRhheag8N4fRjQJWKipIEr1HBxrNyVScyrYPN9PFIsJnZAR6xkakoApN12EnqNzowyRH0nzQo0W6veNlARKzQLPTGYh1W8vE//yeon2a27KwzjRLKTLQ34ikI5Q1gAacsmoFjNDCJXcZEV0TCSh2vSUNyV8/RT9T9pl2zFd3VaKjdKqjhycwhmUwIEraMANNKEFFO7hAZ7g2Zpaj9aL9bocXbNWOyfwA9bbJztsklE=</latexit>
(a) Edge update (b) Node update (c) Global update
Figure 3: Updates in a GN block. Blue indicates the element that is being updated, and black
indicates other elements which are involved in the update (note that the pre-update value of the
blue element is also used in the update). See Equation 1 for details on the notation.
3. φv is applied to each node i, to compute an updated node attribute, vi0 . In our running
example, φv may compute something analogous to the updated position, velocity, and kinetic
energy of each ball. The set of resulting per-node outputs is, V 0 = {vi0 }i=1:N v .
4. ρe→u is applied to E 0 , and aggregates all edge updates, into ē0 , which will then be used in the
next step’s global update. In our running example, ρe→u may compute the summed forces
(which should be zero, in this case, due to Newton’s third law) and the springs’ potential
energies.
5. ρv→u is applied to V 0 , and aggregates all node updates, into v̄0 , which will then be used in
the next step’s global update. In our running example, ρv→u might compute the total kinetic
energy of the system.
6. φu is applied once per graph, and computes an update for the global attribute, u0 . In our
running example, φu might compute something analogous to the net forces and total energy
of the physical system.
Note, though we assume this sequence of steps here, the order is not strictly enforced: it is possible
to reverse the update functions to proceed from global, to per-node, to per-edge updates, for example.
Kearnes et al. (2016) computes edge updates from nodes in a similar manner.
3.2.4 Relational inductive biases in graph networks

Our GN framework imposes several strong relational inductive biases when used as components in
a learning process. First, graphs can express arbitrary relationships among entities, which means
the GN’s input determines how representations interact and are isolated, rather than those choices
being determined by the fixed architecture. For example, the assumption that two entities have a
relationship—and thus should interact—is expressed by an edge between the entities’ corresponding
nodes. Similarly, the absence of an edge expresses the assumption that the nodes have no relationship
and should not influence each other directly.
Second, graphs represent entities and their relations as sets, which are invariant to permutations.
This means GNs are invariant to the order of these elements6 , which is often desirable. For example,
the objects in a scene do not have a natural ordering (see Sec. 2.2).
Third, a GN’s per-edge and per-node functions are reused across all edges and nodes, respectively.
This means GNs automatically support a form of combinatorial generalization (see Section 5.1):
because graphs are composed of edges, nodes, and global features, a single GN can operate on
graphs of different sizes (numbers of edges and nodes) and shapes (edge connectivity).
6
Note, an ordering can be imposed by encoding the indices in the node or edge attributes, or via the edges
themselves (e.g. by encoding a chain or partial ordering).
13
4 Design principles for graph network architectures
The GN framework can be used to implement a wide variety of architectures, in accordance with
the design principles listed above in Section 3.2, which also correspond to the sub-sections (4.1,
4.2, and 4.3) below. In general, the framework is agnostic to specific attribute representations and
functional forms. Here, however, we focus mainly on deep learning architectures, which allow GNs
to act as learnable graph-to-graph function approximators.
4.1 Flexible representations

Graph networks support highly flexible graph representations in two ways: first, in terms of the
representation of the attributes; and second, in terms of the structure of the graph itself.
4.1.1 Attributes
The global, node, and edge attributes of a GN block can use arbitrary representational formats. In
deep learning implementations, real-valued vectors and tensors are most common. However, other
data structures such as sequences, sets, or even graphs could also be used.
The requirements of the problem will often determine what representations should be used for
the attributes. For example, when the input data is an image, the attributes might be represented
as tensors of image patches; however, when the input data is a text document, the attributes might
be sequences of words corresponding to sentences.
For each GN block within a broader architecture, the edge and node outputs typically correspond
to lists of vectors or tensors, one per edge or node, and the global outputs correspond to a single
vector or tensor. This allows a GN’s output to be passed to other deep learning building blocks
such as MLPs, CNNs, and RNNs.
The output of a GN block can also be tailored to the demands of the task. In particular,
◦ An edge-focused GN uses the edges as output, for example to make decisions about interactions
among entities (Kipf et al., 2018; Hamrick et al., 2018).
◦ A node-focused GN uses the nodes as output, for example to reason about physical systems
(Battaglia et al., 2016; Chang et al., 2017; Wang et al., 2018b; Sanchez-Gonzalez et al., 2018).
◦ A graph-focused GN uses the globals as output, for example to predict the potential energy of
a physical system (Battaglia et al., 2016), the properties of a molecule (Gilmer et al., 2017),
or answers to questions about a visual scene (Santoro et al., 2017).
The nodes, edges, and global outputs can also be mixed-and-matched depending on the task. For
example, Hamrick et al. (2018) used both the output edge and global attributes to compute a policy
over actions.
4.1.2 Graph structure

When defining how the input data will be represented as a graph, there are generally two scenarios:
first, the input explicitly specifies the relational structure; and second, the relational structure must
be inferred or assumed. These are not hard distinctions, but extremes along a continuum.
Examples of data with more explicitly specified entities and relations include knowledge graphs,
social networks, parse trees, optimization problems, chemical graphs, road networks, and physical
systems with known interactions. Figures 2a-d illustrate how such data can be expressed as graphs.
Examples of data where the relational structure is not made explicit, and must be inferred or
assumed, include visual scenes, text corpora, programming language source code, and multi-agent
14
systems. In these types of settings, the data may be formatted as a set of entities without relations,
or even just a vector or tensor (e.g., an image). If the entities are not specified explicitly, they might
be assumed, for instance, by treating each word in a sentence (Vaswani et al., 2017) or each local
feature vector in a CNN’s output feature map, as a node (Watters et al., 2017; Santoro et al., 2017;
Wang et al., 2018c) (Figures 2e-f). Or, it might be possible to use a separate learned mechanism to
infer entities from an unstructured signal (Luong et al., 2015; Mnih et al., 2014; Eslami et al., 2016;
van Steenkiste et al., 2018). If relations are not available, the simplest approach is to instantiate all
possible directed edges between entities (Figure 2f). This can be prohibitive for large numbers of
entities, however, because the number of possible edges grows quadratically with the number of
nodes. Thus developing more sophisticated ways of inferring sparse structure from unstructured
data (Kipf et al., 2018) is an important future direction.
4.2 Configurable within-block structure
The structure and functions within a GN block can be configured in different ways, which offers
flexibility in what information is made available as inputs to its functions, as well as how output edge,
node, and global updates are produced. In particular, each φ in Equation 1 must be implemented
with some function, f , where f ’s argument signature determines what information it requires as
input; in Figure 4, the incoming arrows to each φ depict whether u, V , and E are taken as inputs.
Hamrick et al. (2018) and Sanchez-Gonzalez et al. (2018) used the full GN block shown in Figure 4a.
Their φ implementations used neural networks (denoted NNe , NNv , and NNu below, to indicate that
they are different functions with different parameters). Their ρ implementations used elementwise
summation, but averages and max/min could also be used,
φe (ek , vrk , vsk , u) := f e (ek , vrk , vsk , u) = NNe ([ek , vrk , vsk , u]) (2)
φv ē0i , vi , u := f v ē0i , vi , u = NNv [ē0i , vi , u]

φu ē0 , v̄0 , u := f u ē0 , v̄0 , u = NNu [ē0 , v̄0 , u]

X
ρe→v Ei0 := e0k

=
{k: rk =i}
X
0
ρv→u V vi0

:= =
i
X
0
e→u
e0k

ρ E := =
k
where [x, y, z] indicates vector/tensor concatenation. For vector attributes, a MLP is often used for
φ, while for tensors such as image feature maps, CNNs may be more suitable.
The φ functions can also use RNNs, which requires an additional hidden state as input and
output. Figure 4b shows a very simple version of a GN block with RNNs as φ functions: there is no
message-passing in this formulation, and this type of block might be used for recurrent smoothing of
some dynamic graph states. Of course, RNNs as φ functions could also be used in a full GN block
(Figure 4a).
A variety of other architectures can be expressed in the GN framework, often as different
function choices and within-block configurations. The remaining sub-sections explore how a GN’s
within-block structure can be configured in different ways, with examples of published work which
uses such configurations. See the Appendix for details.
15
u, uhid
<latexit sha1_base64="UO6spgXZarocGO1zBKohmEwCj+c=">AAACDHicbZDNSsNAFIVv6l+tf1WXbgaL4EJKIoIuXBTcuKxgW6ENZTKZtENnkjAzEUrIA7jxVdy4UMStD+DOt3GSBtTWAwMf597L3Hu8mDOlbfvLqiwtr6yuVddrG5tb2zv13b2uihJJaIdEPJJ3HlaUs5B2NNOc3sWSYuFx2vMmV3m9d0+lYlF4q6cxdQUehSxgBGtjDeuNgcB67AVpkp2gHx4WKEU6Zn5muuymXQgtglNCA0q1h/XPgR+RRNBQE46V6jt2rN0US80Ip1ltkCgaYzLBI9o3GGJBlZsWx2ToyDg+CiJpXqhR4f6eSLFQaio805nvqOZruflfrZ/o4MJNWRgnmoZk9lGQcKQjlCeDfCYp0XxqABPJzK6IjLHERJv8aiYEZ/7kReieNh276dycNVqXZRxVOIBDOAYHzqEF19CGDhB4gCd4gVfr0Xq23qz3WWvFKmf24Y+sj28GGpw2</latexit>
<latexit sha1_base64="S5XnA5iYIAgqxiI+i0ptSwAiKP4=">AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe5stLAI2FhGMB+QnGFvM5es2ds9dveEEPIfbCwUsfX/2Plv3CRXaOKDgcd7M8zMi1LBjfX9b6+wtr6xuVXcLu3s7u0flA+PmkZlmmGDKaF0O6IGBZfYsNwKbKcaaRIJbEWjm5nfekJtuJL3dpximNCB5DFn1Dqp2U2H/CHrlSt+1Z+DrJIgJxXIUe+Vv7p9xbIEpWWCGtMJ/NSGE6otZwKnpW5mMKVsRAfYcVTSBE04mV87JWdO6ZNYaVfSkrn6e2JCE2PGSeQ6E2qHZtmbif95nczGV+GEyzSzKNliUZwJYhWZvU76XCOzYuwIZZq7WwkbUk2ZdQGVXAjB8surpHlRDfxqcOdXatd5HEU4gVM4hwAuoQa3UIcGMHiEZ3iFN095L96797FoLXj5zDH8gff5A53Djxw=</latexit>
sha1_base64="znt8hwWv6wryqwCugrweUa+jkM8=">AAAB7XicbVA9SwNBEJ3zM8avqGBjsxgEq3Bno4VFwMYygrkEkjPubfaSNXu7y+6eEo78BxsLRWz9P3b+GzcfhSY+GHi8N8PMvFhxZqzvf3tLyyura+uFjeLm1vbObmlvPzQy04TWieRSN2NsKGeC1i2znDaVpjiNOW3Eg6ux33ik2jApbu1Q0SjFPcESRrB1UthWfXaXdUplv+JPgBZJMCPl6mH4dA8AtU7pq92VJEupsIRjY1qBr2yUY20Z4XRUbGeGKkwGuEdbjgqcUhPlk2tH6MQpXZRI7UpYNFF/T+Q4NWaYxq4zxbZv5r2x+J/XymxyEeVMqMxSQaaLkowjK9H4ddRlmhLLh45gopm7FZE+1phYF1DRhRDMv7xIwrNK4FeCG5fGJUxRgCM4hlMI4ByqcA01qAOBB3iGV3jzpPfivXsf09YlbzZzAH/gff4AqE6Qnw==</latexit>
sha1_base64="Nc0DXje6uYB+/0fHlXL99yCq0no=">AAAB7XicbVC7SgNBFL0bXzG+ooKNzWAQrMKujRYWARvLCGYTSNYwO5lNxszODjOzyrLkH2wsFNHS/7HzA/wPJ49CEw9cOJxzL/feE0rOtHHdL6ewtLyyulZcL21sbm3vlHf3fJ2kitAGSXiiWiHWlDNBG4YZTltSURyHnDbD4eXYb95TpVkibkwmaRDjvmARI9hYye/IAbtNu+WKW3UnQIvEm5FK7cB/kNn3e71b/uz0EpLGVBjCsdZtz5UmyLEyjHA6KnVSTSUmQ9ynbUsFjqkO8sm1I3RslR6KEmVLGDRRf0/kONY6i0PbGWMz0PPeWPzPa6cmOg9yJmRqqCDTRVHKkUnQ+HXUY4oSwzNLMFHM3orIACtMjA2oZEPw5l9eJP5p1XOr3rVN4wKmKMIhHMEJeHAGNbiCOjSAwB08wjO8OInz5Lw6b9PWgjOb2Yc/cD5+AGXTkqw=</latexit>
u u0 , u0hid
<latexit sha1_base64="W2hu4ghYojV0SmOcBbcdBDITOSo=">AAACDnicbVC7TsMwFHV4lvIKMLJYVBUMqEoQEgwMlVgYi0QfUhtFjuu0Vm0nsh2kKsoXsPArLAwgxMrMxt/gpBlKy5EsHZ9zr+69J4gZVdpxfqyV1bX1jc3KVnV7Z3dv3z447KgokZi0ccQi2QuQIowK0tZUM9KLJUE8YKQbTG5zv/tIpKKReNDTmHgcjQQNKUbaSL5dH3Ckx0GYJtnpOZz7+AWXPB3TYebbNafhFIDLxC1JDZRo+fb3YBjhhBOhMUNK9V0n1l6KpKaYkaw6SBSJEZ6gEekbKhAnykuLczJYN8oQhpE0T2hYqPMdKeJKTXlgKvMd1aKXi/95/USH115KRZxoIvBsUJgwqCOYZwOHVBKs2dQQhCU1u0I8RhJhbRKsmhDcxZOXSeei4ToN9/6y1rwp46iAY3ACzoALrkAT3IEWaAMMnsALeAPv1rP1an1Yn7PSFavsOQJ/YH39AtnQnJg=</latexit>
V, Vhid
<latexit sha1_base64="brwuxF74R6OEOykh378as/RBfzE=">AAAB+nicbVBNS8NAFHzxs9avVI9eFovgQUoigh48FLx4rGDTQhvCZrNtl+4mYXejlNif4sWDIl79Jd78N27aHLR1YGGYeY83O2HKmdKO822trK6tb2xWtqrbO7t7+3btwFNJJgltk4QnshtiRTmLaVszzWk3lRSLkNNOOL4p/M4DlYol8b2epNQXeBizASNYGymwa94Z8oK+wHokRT5i0TSw607DmQEtE7ckdSjRCuyvfpSQTNBYE46V6rlOqv0cS80Ip9NqP1M0xWSMh7RnaIwFVX4+iz5FJ0aJ0CCR5sUazdTfGzkWSk1EaCaLjGrRK8T/vF6mB1d+zuI00zQm80ODjCOdoKIHFDFJieYTQzCRzGRFZIQlJtq0VTUluItfXibeecN1Gu7dRb15XdZRgSM4hlNw4RKacAstaAOBR3iGV3iznqwX6936mI+uWOXOIfyB9fkD20qTuA==</latexit>
<latexit sha1_base64="Fc8T4ygtia14k1z/CDji4ezWDqY=">AAAB7XicbVA9SwNBEJ3zM8avqKXNYhCswp2NFhYBG8sI5gOSM+xtNsmavd1jdy4QjvwHGwtFbP0/dv4bN8kVmvhg4PHeDDPzokQKi77/7a2tb2xubRd2irt7+weHpaPjhtWpYbzOtNSmFVHLpVC8jgIlbyWG0ziSvBmNbmd+c8yNFVo94CThYUwHSvQFo+ikRicZisdxt1T2K/4cZJUEOSlDjlq39NXpaZbGXCGT1Np24CcYZtSgYJJPi53U8oSyER3wtqOKxtyG2fzaKTl3So/0tXGlkMzV3xMZja2dxJHrjCkO7bI3E//z2in2r8NMqCRFrthiUT+VBDWZvU56wnCGcuIIZUa4WwkbUkMZuoCKLoRg+eVV0risBH4luPfL1Zs8jgKcwhlcQABXUIU7qEEdGDzBM7zCm6e9F+/d+1i0rnn5zAn8gff5A59Hjx0=</latexit>
sha1_base64="HCiXjOq04H3f4Ed7vqyiRfd+2dI=">AAAB7XicbVA9SwNBEJ2LXzF+nQo2NotBsAp3NlpYBGwsI5hLIDnj3mYvWbO3e+zuRcKR/2BjoYit/8fOf+Pmo9DEBwOP92aYmRelnGnjed9OYWV1bX2juFna2t7Z3XP3DwItM0VonUguVTPCmnImaN0ww2kzVRQnEaeNaHA98RtDqjST4s6MUhomuCdYzAg2VgraaZ/dDztu2at4U6Bl4s9JuXoUPD0AQK3jfrW7kmQJFYZwrHXL91IT5lgZRjgdl9qZpikmA9yjLUsFTqgO8+m1Y3RqlS6KpbIlDJqqvydynGg9SiLbmWDT14veRPzPa2UmvgxzJtLMUEFmi+KMIyPR5HXUZYoSw0eWYKKYvRWRPlaYGBtQyYbgL768TILziu9V/FubxhXMUIRjOIEz8OECqnADNagDgUd4hld4c6Tz4rw7H7PWgjOfOYQ/cD5/AKnSkKA=</latexit>
sha1_base64="/voSHBGyFE5xYPVKPYM/GTIarLQ=">AAAB7XicbVC7SgNBFL3rM8ZXVLCxWQyCVdi10cIiYGMZwWwCyRpmJ7PJmNmZYWY2siz5BxsLRbT0f+z8AP/DyaPQxAMXDufcy733RJJRbTzvy1laXlldWy9sFDe3tnd2S3v7gRapwqSOBROqGSFNGOWkbqhhpCkVQUnESCMaXI39xpAoTQW/NZkkYYJ6nMYUI2OloC379G7YKZW9ijeBu0j8GSlXD4MHmX2/1zqlz3ZX4DQh3GCGtG75njRhjpShmJFRsZ1qIhEeoB5pWcpRQnSYT64duSdW6bqxULa4cSfq74kcJVpnSWQ7E2T6et4bi/95rdTEF2FOuUwN4Xi6KE6Za4Q7ft3tUkWwYZklCCtqb3VxHymEjQ2oaEPw519eJMFZxfcq/o1N4xKmKMARHMMp+HAOVbiGGtQBwz08wjO8OMJ5cl6dt2nrkjObOYA/cD5+AGdXkq0=</latexit>
v V 0 , Vhid
0
<latexit sha1_base64="WhBBytL7AOLCOPFKqkdbfTZ2xq4=">AAAB/HicbVBPS8MwHE3nvzn/VXf0EhwyDzJaEfTgYeDF4wTXDbZS0jTdwpK0JKkwyvwqXjwo4tUP4s1vY7r1oJsPAo/3fj9+Ly9MGVXacb6tytr6xuZWdbu2s7u3f2AfHnkqySQmXZywRPZDpAijgnQ11Yz0U0kQDxnphZPbwu89EqloIh70NCU+RyNBY4qRNlJg173mOfSawZAjPZY8H9NoFtgNp+XMAVeJW5IGKNEJ7K9hlOCME6ExQ0oNXCfVfo6kppiRWW2YKZIiPEEjMjBUIE6Un8/Dz+CpUSIYJ9I8oeFc/b2RI67UlIdmssiolr1C/M8bZDq+9nMq0kwTgReH4oxBncCiCRhRSbBmU0MQltRkhXiMJMLa9FUzJbjLX14l3kXLdVru/WWjfVPWUQXH4AScARdcgTa4Ax3QBRhMwTN4BW/Wk/VivVsfi9GKVe7UwR9Ynz+huZQa</latexit>
<latexisha1_be64="iqmo18xzMbJ35wZs9C4FNu1joOA0ltdTSUcZqI95s7n2B3DP8rAogTli6MaYc=">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3PZn<g/zQl9dawtXseIx8iAG<>k/lQatwe=xi</>latexi>
e
E, Ehid
<latexit sha1_base64="DYZek5SmevKS8py25dx0aIqUbBY=">AAAB+nicbVDLSsNAFL3xWesr1aWbwSK4kJKIoAsXBSm4rGAf0IYwmUzboTNJmJkoJfZT3LhQxK1f4s6/cdJmoa0HBg7n3Ms9c4KEM6Ud59taWV1b39gsbZW3d3b39u3KQVvFqSS0RWIey26AFeUsoi3NNKfdRFIsAk47wfgm9zsPVCoWR/d6klBP4GHEBoxgbSTfrjTOUMPvC6xHUmQjFk59u+rUnBnQMnELUoUCTd/+6ocxSQWNNOFYqZ7rJNrLsNSMcDot91NFE0zGeEh7hkZYUOVls+hTdGKUEA1iaV6k0Uz9vZFhodREBGYyz6gWvVz8z+ulenDlZSxKUk0jMj80SDnSMcp7QCGTlGg+MQQTyUxWREZYYqJNW2VTgrv45WXSPq+5Ts29u6jWr4s6SnAEx3AKLlxCHW6hCS0g8AjP8Apv1pP1Yr1bH/PRFavYOYQ/sD5/AKXVk5Y=</latexit>
<latexit sha1_base64="pLq6KB/1S9uyUeWp/G4byg43mK0=">AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe5stLAI2FhGMB+QnGFvM5es2ds9dveEEPIfbCwUsfX/2Plv3CRXaOKDgcd7M8zMi1LBjfX9b6+wtr6xuVXcLu3s7u0flA+PmkZlmmGDKaF0O6IGBZfYsNwKbKcaaRIJbEWjm5nfekJtuJL3dpximNCB5DFn1Dqp2U2H/AF75Ypf9ecgqyTISQVy1Hvlr25fsSxBaZmgxnQCP7XhhGrLmcBpqZsZTCkb0QF2HJU0QRNO5tdOyZlT+iRW2pW0ZK7+npjQxJhxErnOhNqhWfZm4n9eJ7PxVTjhMs0sSrZYFGeCWEVmr5M+18isGDtCmebuVsKGVFNmXUAlF0Kw/PIqaV5UA78a3PmV2nUeRxFO4BTOIYBLqMEt1KEBDB7hGV7hzVPei/fufSxaC14+cwx/4H3+AIWDjww=</latexit>
sha1_base64="gRKFy+QFytmwqWy0cvo5FmmPz8I=">AAAB7XicbVA9SwNBEJ3zM8avqGBjsxgEq3Bno4VFwMYygrkEkjPubeaSNXu3x+6eEo78BxsLRWz9P3b+GzcfhSY+GHi8N8PMvDAVXBvX/XaWlldW19YLG8XNre2d3dLevq9lphjWmRRSNUOqUfAE64Ybgc1UIY1DgY1wcDX2G4+oNJfJrRmmGMS0l/CIM2qs5LfTPr/DTqnsVtwJyCLxZqRcPfSf7gGg1il9tbuSZTEmhgmqdctzUxPkVBnOBI6K7UxjStmA9rBlaUJj1EE+uXZETqzSJZFUthJDJurviZzGWg/j0HbG1PT1vDcW//NamYkugpwnaWYwYdNFUSaIkWT8OulyhcyIoSWUKW5vJaxPFWXGBlS0IXjzLy8S/6ziuRXvxqZxCVMU4AiO4RQ8OIcqXEMN6sDgAZ7hFd4c6bw4787HtHXJmc0cwB84nz+QDpCP</latexit>
sha1_base64="74MJShuZzxGyM2nLY3EY87InuhI=">AAAB7XicbVC7SgNBFJ2NrxhfUcHGZjAIVmHXRguLgI1lBLMJJGuYndxNxszODjOzyrLkH2wsFNHS/7HzA/wPJ49CEw9cOJxzL/feE0rOtHHdL6ewtLyyulZcL21sbm3vlHf3fJ2kikKDJjxRrZBo4ExAwzDDoSUVkDjk0AyHl2O/eQ9Ks0TcmExCEJO+YBGjxFjJ78gBu4VuueJW3QnwIvFmpFI78B9k9v1e75Y/O72EpjEIQznRuu250gQ5UYZRDqNSJ9UgCR2SPrQtFSQGHeSTa0f42Co9HCXKljB4ov6eyEmsdRaHtjMmZqDnvbH4n9dOTXQe5EzI1ICg00VRyrFJ8Ph13GMKqOGZJYQqZm/FdEAUocYGVLIhePMvLxL/tOq5Ve/apnGBpiiiQ3SETpCHzlANXaE6aiCK7tAjekYvTuI8Oa/O27S14Mxm9tEfOB8/TZOSnA==</latexit>
E 0 , Ehid
0
<latexit sha1_base64="ZYHbDTDbKJNm66GNrt+dUMK6558=">AAAB/HicbVDLSgMxFM3UV62v0S7dBIvUhZQZEXThoiAFlxXsA9phyGQybWiSGZKMUIb6K25cKOLWD3Hn35hpZ6GtBwKHc+7lnpwgYVRpx/m2SmvrG5tb5e3Kzu7e/oF9eNRVcSox6eCYxbIfIEUYFaSjqWakn0iCeMBIL5jc5n7vkUhFY/GgpwnxOBoJGlGMtJF8u9qqn8NW3R9ypMeSZ2Mazny75jScOeAqcQtSAwXavv01DGOcciI0Zkipgesk2suQ1BQzMqsMU0UShCdoRAaGCsSJ8rJ5+Bk8NUoIo1iaJzScq783MsSVmvLATOYZ1bKXi/95g1RH115GRZJqIvDiUJQyqGOYNwFDKgnWbGoIwpKarBCPkURYm74qpgR3+curpHvRcJ2Ge39Za94UdZTBMTgBZ8AFV6AJ7kAbdAAGU/AMXsGb9WS9WO/Wx2K0ZBU7VfAH1ucPbBGT+A==</latexit>
Edge block Node block Global block
(a) Full GN block (b) Independent recurrent block
<latexit sha1_base64="S5XnA5iYIAgqxiI+i0ptSwAiKP4=">AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe5stLAI2FhGMB+QnGFvM5es2ds9dveEEPIfbCwUsfX/2Plv3CRXaOKDgcd7M8zMi1LBjfX9b6+wtr6xuVXcLu3s7u0flA+PmkZlmmGDKaF0O6IGBZfYsNwKbKcaaRIJbEWjm5nfekJtuJL3dpximNCB5DFn1Dqp2U2H/CHrlSt+1Z+DrJIgJxXIUe+Vv7p9xbIEpWWCGtMJ/NSGE6otZwKnpW5mMKVsRAfYcVTSBE04mV87JWdO6ZNYaVfSkrn6e2JCE2PGSeQ6E2qHZtmbif95nczGV+GEyzSzKNliUZwJYhWZvU76XCOzYuwIZZq7WwkbUk2ZdQGVXAjB8surpHlRDfxqcOdXatd5HEU4gVM4hwAuoQa3UIcGMHiEZ3iFN095L96797FoLXj5zDH8gff5A53Djxw=</latexit>
sha1_base64="znt8hwWv6wryqwCugrweUa+jkM8=">AAAB7XicbVA9SwNBEJ3zM8avqGBjsxgEq3Bno4VFwMYygrkEkjPubfaSNXu7y+6eEo78BxsLRWz9P3b+GzcfhSY+GHi8N8PMvFhxZqzvf3tLyyura+uFjeLm1vbObmlvPzQy04TWieRSN2NsKGeC1i2znDaVpjiNOW3Eg6ux33ik2jApbu1Q0SjFPcESRrB1UthWfXaXdUplv+JPgBZJMCPl6mH4dA8AtU7pq92VJEupsIRjY1qBr2yUY20Z4XRUbGeGKkwGuEdbjgqcUhPlk2tH6MQpXZRI7UpYNFF/T+Q4NWaYxq4zxbZv5r2x+J/XymxyEeVMqMxSQaaLkowjK9H4ddRlmhLLh45gopm7FZE+1phYF1DRhRDMv7xIwrNK4FeCG5fGJUxRgCM4hlMI4ByqcA01qAOBB3iGV3jzpPfivXsf09YlbzZzAH/gff4AqE6Qnw==</latexit>
sha1_base64="Nc0DXje6uYB+/0fHlXL99yCq0no=">AAAB7XicbVC7SgNBFL0bXzG+ooKNzWAQrMKujRYWARvLCGYTSNYwO5lNxszODjOzyrLkH2wsFNHS/7HzA/wPJ49CEw9cOJxzL/feE0rOtHHdL6ewtLyyulZcL21sbm3vlHf3fJ2kitAGSXiiWiHWlDNBG4YZTltSURyHnDbD4eXYb95TpVkibkwmaRDjvmARI9hYye/IAbtNu+WKW3UnQIvEm5FK7cB/kNn3e71b/uz0EpLGVBjCsdZtz5UmyLEyjHA6KnVSTSUmQ9ynbUsFjqkO8sm1I3RslR6KEmVLGDRRf0/kONY6i0PbGWMz0PPeWPzPa6cmOg9yJmRqqCDTRVHKkUnQ+HXUY4oSwzNLMFHM3orIACtMjA2oZEPw5l9eJP5p1XOr3rVN4wKmKMIhHMEJeHAGNbiCOjSAwB08wjO8OInz5Lw6b9PWgjOb2Yc/cD5+AGXTkqw=</latexit>
u u0
<latexit sha1_base64="Z/n1gIms2/ONBt0R58c8NGdBbqU=">AAAB8nicbVDLSsNAFL2pr1pfVZduBovoqiQi6MJFwY3LCvYBbSiT6aQdOpmEmRuhhH6GGxeKuPVr3Pk3TtostPXAwOGce5lzT5BIYdB1v53S2vrG5lZ5u7Kzu7d/UD08aps41Yy3WCxj3Q2o4VIo3kKBkncTzWkUSN4JJne533ni2ohYPeI04X5ER0qEglG0Uq8fURwHYZbOzgfVmlt35yCrxCtIDQo0B9Wv/jBmacQVMkmN6Xlugn5GNQom+azSTw1PKJvQEe9ZqmjEjZ/NI8/ImVWGJIy1fQrJXP29kdHImGkU2Mk8oln2cvE/r5dieONnQiUpcsUWH4WpJBiT/H4yFJozlFNLKNPCZiVsTDVlaFuq2BK85ZNXSfuy7rl17+Gq1rgt6ijDCZzCBXhwDQ24hya0gEEMz/AKbw46L86787EYLTnFzjH8gfP5A1s4kUQ=</latexit>
v!u
⇢
<latexit sha1_base64="8QVocR3pGD0i/QL+G1OhPZDl9fE=">AAAB/nicbVBNS8NAEN3Ur1q/ouLJy2IRPJVEBD0WvXisYD+giWWz3TRLN9mwO6mUUPCvePGgiFd/hzf/jds2B219MPB4b4aZeUEquAbH+bZKK6tr6xvlzcrW9s7unr1/0NIyU5Q1qRRSdQKimeAJawIHwTqpYiQOBGsHw5up3x4xpblM7mGcMj8mg4SHnBIwUs8+8lQkH/KRp/ggAqKUfMTZpGdXnZozA14mbkGqqECjZ395fUmzmCVABdG66zop+DlRwKlgk4qXaZYSOiQD1jU0ITHTfj47f4JPjdLHoVSmEsAz9fdETmKtx3FgOmMCkV70puJ/XjeD8MrPeZJmwBI6XxRmAoPE0yxwnytGQYwNIVRxcyumEVGEgkmsYkJwF19eJq3zmuvU3LuLav26iKOMjtEJOkMuukR1dIsaqIkoytEzekVv1pP1Yr1bH/PWklXMHKI/sD5/AAdVlig=</latexit>
v v
V
<latexit sha1_base64="xc4uzoZiBSxZUZkArgltxczS6nM=">AAAB6HicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiIVFwMYyAfMByRH2NnPJmr29Y3dPCEd+gY2FIrb+JDv/jZvkCk18MPB4b4aZeUEiuDau++0UNja3tneKu6W9/YPDo/LxSVvHqWLYYrGIVTegGgWX2DLcCOwmCmkUCOwEk7u533lCpXksH8w0QT+iI8lDzqixUrM9KFfcqrsAWSdeTiqQozEof/WHMUsjlIYJqnXPcxPjZ1QZzgTOSv1UY0LZhI6wZ6mkEWo/Wxw6IxdWGZIwVrakIQv190RGI62nUWA7I2rGetWbi/95vdSEN37GZZIalGy5KEwFMTGZf02GXCEzYmoJZYrbWwkbU0WZsdmUbAje6svrpH1V9dyq17yu1G/zOIpwBudwCR7UoA730IAWMEB4hld4cx6dF+fd+Vi2Fpx85hT+wPn8AbDjjNQ=</latexit>
<latexit
V0
<latexit sha1_base64="gAQ7qdt3IKvK5oBqK3uN1PHYi1k=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNoFe4koIVFwMYyivmA5Ah7m7lkyd7esbsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCARXBvX/XYKa+sbm1vF7dLO7t7+QfnwqKXjVDFssljEqhNQjYJLbBpuBHYShTQKBLaD8e3Mbz+h0jyWj2aSoB/RoeQhZ9RY6aF13i9X3Ko7B1klXk4qkKPRL3/1BjFLI5SGCap113MT42dUGc4ETku9VGNC2ZgOsWuppBFqP5tfOiVnVhmQMFa2pCFz9fdERiOtJ1FgOyNqRnrZm4n/ed3UhNd+xmWSGpRssShMBTExmb1NBlwhM2JiCWWK21sJG1FFmbHhlGwI3vLLq6R1WfXcqndfq9Rv8jiKcAKncAEeXEEd7qABTWAQwjO8wpszdl6cd+dj0Vpw8plj+APn8wcRSI0F</latexit>
V <latexisha1_be64="iqmo18xzMbJ35wZs9C4FNu1joOA0ltdTSUcZqI95s7n2B3DP8rAogTli6MaYc=">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3PZn<g/zQl9dawtXseIx8iAG<>k/lQatwe=xi</>latexi>
V0
e!v e!v
⇢
<latexit sha1_base64="s3/Cw/iD/Ic9TAit26LWmPV1hK0=">AAAB/nicbVBNS8NAEN3Ur1q/ouLJy2IRPJVEBD0WvXisYD+giWWznTRLN9mwu6mUUPCvePGgiFd/hzf/jds2B219MPB4b4aZeUHKmdKO822VVlbX1jfKm5Wt7Z3dPXv/oKVEJik0qeBCdgKigLMEmpppDp1UAokDDu1geDP12yOQionkXo9T8GMySFjIKNFG6tlHnozEQw6eZINIEynFIx5NenbVqTkz4GXiFqSKCjR69pfXFzSLIdGUE6W6rpNqPydSM8phUvEyBSmhQzKArqEJiUH5+ez8CT41Sh+HQppKNJ6pvydyEis1jgPTGRMdqUVvKv7ndTMdXvk5S9JMQ0Lni8KMYy3wNAvcZxKo5mNDCJXM3IppRCSh2iRWMSG4iy8vk9Z5zXVq7t1FtX5dxFFGx+gEnSEXXaI6ukUN1EQU5egZvaI368l6sd6tj3lrySpmDtEfWJ8/7hmWGA==</latexit>
⇢
e e
E
<latexit sha1_base64="iJ/x8cSgmmYNbMN8WtCvsNrlH/U=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0IOHgggeW7Af0Iay2U7atZtN2N0IJfQXePGgiFd/kjf/jds2B219MPB4b4aZeUEiuDau++0U1tY3NreK26Wd3b39g/LhUUvHqWLYZLGIVSegGgWX2DTcCOwkCmkUCGwH49uZ335CpXksH8wkQT+iQ8lDzqixUuOuX664VXcOskq8nFQgR71f/uoNYpZGKA0TVOuu5ybGz6gynAmclnqpxoSyMR1i11JJI9R+Nj90Ss6sMiBhrGxJQ+bq74mMRlpPosB2RtSM9LI3E//zuqkJr/2MyyQ1KNliUZgKYmIy+5oMuEJmxMQSyhS3txI2oooyY7Mp2RC85ZdXSeui6rlVr3FZqd3kcRThBE7hHDy4ghrcQx2awADhGV7hzXl0Xpx352PRWnDymWP4A+fzB5cfjMM=</latexit>
<latexit
E <latexit sha1_base64="gRKFy+QFytmwqWy0cvo5FmmPz8I=">AAAB7XicbVA9SwNBEJ3zM8avq
Edge block Node block Global block Edge block Node block Global block
(c) Message-passing neural network (d) Non- oca neura network
u u u u u
v!u
⇢
v
V V
⇢e!u
e
E
Edge b ock Node b ock G oba b ock Edge b ock Node b ock G oba b ock
(e) Re at on network (f) Deep set
Figure 4: Different internal GN block configurations. See Section 3.2 for details on the notation,
and Section 4 for details about each variant. (a) A full GN predicts node, edge, and global output
attributes based on incoming node, edge, and global attributes. (b) An independent, recurrent
update block takes input and hidden graphs, and the φ functions are RNNs (Sanchez-Gonzalez
et a 2018) (c) An MPNN (G mer et a 2017) pred cts node edge and g oba output attr butes
based on incoming node, edge, and global attributes. Note that the global prediction does not
include aggregated edges. (d) A NLNN (Wang et al., 2018c) only predicts node output attributes.
(e) A relation network (Raposo et al., 2017; Santoro et al., 2017) only uses the edge predictions
to predict global attributes. (f) A Deep Set (Zaheer et al., 2017) bypasses the edge update and
predicts updated global attributes.
4.2.1 Message-passing neural network (MPNN)

Gilmer et al. (2017)’s MPNN generalizes a number of previous architectures and can be translated
naturally into the GN formalism. Following the MPNN paper’s terminology (see Gilmer et al.
(2017), pages 2-4):
◦ the message function, Mt , plays the role of the GN’s φe , but does not take u as input,
◦ elementwise summation is used for the GN’s ρe→v ,
◦ the update function, Ut , plays the role of the GN’s φv ,
16
v
V
<latexit sha1_base64="xc4uzoZiBSxZUZkArgltxczS6nM=">AAAB6HicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiIVFwMYyAfMByRH2NnPJmr29Y3dPCEd+gY2FIrb+JDv/jZvkCk18MPB4b4aZeUEiuDau++0UNja3tneKu6W9/YPDo/LxSVvHqWLYYrGIVTegGgWX2DLcCOwmCmkUCOwEk7u533lCpXksH8w0QT+iI8lDzqixUrM9KFfcqrsAWSdeTiqQozEof/WHMUsjlIYJqnXPcxPjZ1QZzgTOSv1UY0LZhI6wZ6mkEWo/Wxw6IxdWGZIwVrakIQv190RGI62nUWA7I2rGetWbi/95vdSEN37GZZIalGy5KEwFMTGZf02GXCEzYmoJZYrbWwkbU0WZsdmUbAje6svrpH1V9dyq17yu1G/zOIpwBudwCR7UoA730IAWMEB4hld4cx6dF+fd+Vi2Fpx85hT+wPn8AbDjjNQ=</latexit>
<latexit
V0
⇢e!v
↵e
<latexit sha1_base64="mjs48M94BmT64MLe6tKpQUPkNJM=">AAAB73icbVA9SwNBEJ3zM8avqKXNYRCswp0IWlgEbCwjmA9IzjC3mUuW7O2du3tCCPkTNhaK2Pp37Pw3bpIrNPHBwOO9GWbmhang2njet7Oyura+sVnYKm7v7O7tlw4OGzrJFKM6S0SiWiFqElxS3XAjqJUqwjgU1AyHN1O/+URK80Tem1FKQYx9ySPO0Fip1UGRDvCBuqWyV/FmcJeJn5My5Kh1S1+dXsKymKRhArVu+15qgjEqw5mgSbGTaUqRDbFPbUslxqSD8ezeiXtqlZ4bJcqWNO5M/T0xxljrURzazhjNQC96U/E/r52Z6CoYc5lmhiSbL4oy4ZrEnT7v9rgiZsTIEmSK21tdNkCFzNiIijYEf/HlZdI4r/hexb+7KFev8zgKcAwncAY+XEIVbqEGdWAg4Ble4c15dF6cd+dj3rri5DNH8AfO5w8AXY/p</latexit>
e
<latexit sha1_base64="EsSika+7GeuHZg/021QVL5D37Tg=">AAAB7nicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoAcPAS8eI5gHJGuYnfQmQ2YfzPQKIeQjvHhQxKvf482/cZLsQRMLGoqqbrq7glRJQ6777RTW1jc2t4rbpZ3dvf2D8uFR0ySZFtgQiUp0O+AGlYyxQZIUtlONPAoUtoLR7cxvPaE2MokfaJyiH/FBLEMpOFmp1Q2Q+CP2yhW36s7BVomXkwrkqPfKX91+IrIIYxKKG9Px3JT8CdckhcJpqZsZTLkY8QF2LI15hMafzM+dsjOr9FmYaFsxsbn6e2LCI2PGUWA7I05Ds+zNxP+8TkbhtT+RcZoRxmKxKMwUo4TNfmd9qVGQGlvChZb2ViaGXHNBNqGSDcFbfnmVNC+qnlv17i8rtZs8jiKcwCmcgwdXUIM7qEMDBIzgGV7hzUmdF+fd+Vi0Fpx85hj+wPn8ATfaj3U=</latexit>
e
Figure 5: NLNNs as GNs. A schematic showing how NLNNs (Wang et al., 2018c) are implemented
by the φe and ρe→v under the GN framework. Typically, NLNNs assume that different regions of
an image (or words in a sentence) correspond to nodes in a fully connected graph, and the attention
mechanism defines a weighted sum over nodes during the aggregation step.
◦ the readout function, R, plays the role of the GN’s φu , but does not take u or E 0 as input,
and thus an analog to the GN’s ρe→u is not required;
◦ dmaster serves a roughly similar purpose to the GN’s u, but is defined as an extra node
connected to all others, and thus does not influence the edge and global updates directly. It
can then be represented in the GN’s V .
Figure 4c shows how an MPNN is structured, according to the GN framework. For details and
various MPNN architectures, see the Appendix.
4.2.2 Non-local neural networks (NLNN)

Wang et al. (2018c)’s NLNN, which unifies various “intra-/self-/vertex-/graph-attention” approaches
(Lin et al., 2017; Vaswani et al., 2017; Hoshen, 2017; Veličković et al., 2018; Shaw et al., 2018),
can also be translated into the GN formalism. The label “attention” refers to how the nodes are
updated: each node update is based on a weighted sum of (some function of) the node attributes of
its neighbors, where the weight between a node and one of its neighbors is computed by a scalar
pairwise function between their attributes (and then normalized across neighbors). The published
NLNN formalism does not explicitly include edges, and instead computes pairwise attention weights
between all nodes. But various NLNN-compliant models, such as the vertex attention interaction
network (Hoshen, 2017) and graph attention network (Veličković et al., 2018), are able to handle
explicit edges by effectively setting to zero the weights between nodes which do not share an edge.
As Figures 4d and 5 illustrate, the φe is factored into the scalar pairwise-interaction function
which returns the unnormalized attention term, denoted αe (vrk , vsk ) = a0k , and a vector-valued
non-pairwise term, denoted β e (vsk ) = b0k . In the ρe→v aggregation, the a0k terms are normalized
across each receiver’s edges, b0k , and elementwise summed:
φe (ek , vrk , vsk , u) := f e (vrk , vsk ) = (αe (vrk , vsk ) , β e (vsk )) = (a0k , b0k ) = e0k
φv ē0i , vi , u := f v (ē0i )

1 X
ρe→v Ei0 := P a0k b0k

0
{k: rk =i} ak {k: rk =i}
In the NLNN paper’s terminology (see Wang et al. (2018c), pages 2-4):
◦ their f plays the role of the above α,
17
◦ their g plays the role of the above β.
This formulation may be helpful for focusing only on those interactions which are most relevant for
the downstream task, especially when the input entities were a set, from which a graph was formed
by adding all possible edges between them.
Vaswani et al. (2017)’s multi-headed self-attention mechanism adds an interesting feature, where
the φe and ρe→v are implemented by a parallel set of functions, whose results are concatenated
together as the final step of ρe→v . This can be interpreted as using typed edges, where the different
types index into different φe component functions, analogous to Li et al. (2016).
For details and various NLNN architectures, see the Appendix.
4.2.3 Other graph network variants

The full GN (Equation 2) can be used to predict a full graph, or any subset of (u0 , V 0 , E 0 ), as
outlined in Section 4.1.1. For example, to predict a global property of a graph, V 0 and E 0 can just
be ignored. Similarly, if global, node, or edge attributes are unspecified in the inputs, those vectors
can be zero-length, i.e., not taken as explicit input arguments. The same idea applies for other GN
variants which do not use the full set of mapping (φ) and reduction (ρ) functions. For instance,
Interaction Networks (Battaglia et al., 2016; Watters et al., 2017) and the Neural Physics Engine
(Chang et al., 2017) use a full GN but for the absence of the global to update the edge properties
(see Appendix for details).
Various models, including CommNet (Sukhbaatar et al., 2016), structure2vec (Dai et al., 2016)
(in the version of (Dai et al., 2017)), and Gated Graph Sequence Neural Networks (Li et al., 2016)
have used a φe which does not directly compute pairwise interactions, but instead ignore the receiver
node, operating only on the sender node and in some cases an edge attribute. This can be expressed
by implementations of φe with the following signatures, such as:
φe (ek , vrk , vsk , u) := f e (vsk )
or φe (ek , vrk , vsk , u) := vsk + f e (ek )
or φe (ek , vrk , vsk , u) := f e (ek , vsk ) .
See the Appendix for further details.
Relation Networks (Raposo et al., 2017; Santoro et al., 2017) bypass the node update entirely
and predict the global output from pooled edge information directly (see also Figure 4e),
φe (ek , vrk , vsk , u) := f e (vrk , vsk ) = NNe ([vrk , vsk ])

φu ē0 , v̄0 , u := f u ē0 = NNu ē0

X
ρe→u E 0 := e0k

=
k
Deep Sets (Zaheer et al., 2017) bypass the edges update completely and predict the global output
from pooled nodes information directly (Figure 4f),
φv (ēi , vi , u) := f v (vi , u) = NNv ([vi , u])

φu ē0 , v̄0 , u := f u v̄0 = NNu v̄0

X
ρv→u V 0 := vi0

=
i
PointNet (Qi et al., 2017) use similar update rule, with a max-aggregation for ρv→u and a two-step
node update.
18
G0
<latexit sha1_base64="vj48jMMQe2f55rU6zb6RZp+K9y4=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiIVFwELLiOYDkiPsbfaSJXt7x+6cEI78BBsLRWz9RXb+GzfJFZr4YODx3gwz84JECoOu++0U1tY3NreK26Wd3b39g/LhUcvEqWa8yWIZ605ADZdC8SYKlLyTaE6jQPJ2ML6Z+e0nro2I1SNOEu5HdKhEKBhFKz3c9t1+ueJW3TnIKvFyUoEcjX75qzeIWRpxhUxSY7qem6CfUY2CST4t9VLDE8rGdMi7lioaceNn81On5MwqAxLG2pZCMld/T2Q0MmYSBbYzojgyy95M/M/rphhe+ZlQSYpcscWiMJUEYzL7mwyE5gzlxBLKtLC3EjaimjK06ZRsCN7yy6ukdVH13Kp3f1mpX+dxFOEETuEcPKhBHe6gAU1gMIRneIU3RzovzrvzsWgtOPnMMfyB8/kDvamNaA==</latexit>
GN1
<latexit sha1_base64="oAmr7/S238q10w2wEvXkfEGmAr8=">AAAB9HicbVDLSgMxFL1TX7W+qi7dBIvgqsyIUBcuCi50JRXsA9qhZNJMG5pkxiRTKEO/w40LRdz6Me78GzPtLLT1QOBwzr3ckxPEnGnjut9OYW19Y3OruF3a2d3bPygfHrV0lChCmyTikeoEWFPOJG0aZjjtxIpiEXDaDsY3md+eUKVZJB/NNKa+wEPJQkawsZLfE9iMlEhv72d9r1+uuFV3DrRKvJxUIEejX/7qDSKSCCoN4VjrrufGxk+xMoxwOiv1Ek1jTMZ4SLuWSiyo9tN56Bk6s8oAhZGyTxo0V39vpFhoPRWBncxC6mUvE//zuokJr/yUyTgxVJLFoTDhyEQoawANmKLE8KklmChmsyIywgoTY3sq2RK85S+vktZF1XOr3sNlpX6d11GEEziFc/CgBnW4gwY0gcATPMMrvDkT58V5dz4WowUn3zmGP3A+fwCg3ZH4</latexit>
G1
<latexit sha1_base64="YNShseMoKm2HdChKvcjMRmoBu5o=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiIVFwELLiOYDkiPsbfaSJXt7x+6cEI78BBsLRWz9RXb+GzfJFZr4YODx3gwz84JECoOu++0U1tY3NreK26Wd3b39g/LhUcvEqWa8yWIZ605ADZdC8SYKlLyTaE6jQPJ2ML6Z+e0nro2I1SNOEu5HdKhEKBhFKz3c9r1+ueJW3TnIKvFyUoEcjX75qzeIWRpxhUxSY7qem6CfUY2CST4t9VLDE8rGdMi7lioaceNn81On5MwqAxLG2pZCMld/T2Q0MmYSBbYzojgyy95M/M/rphhe+ZlQSYpcscWiMJUEYzL7mwyE5gzlxBLKtLC3EjaimjK06ZRsCN7yy6ukdVH13Kp3f1mpX+dxFOEETuEcPKhBHe6gAU1gMIRneIU3RzovzrvzsWgtOPnMMfyB8/kDvy2NaQ==</latexit>
GN2
<latexit sha1_base64="pet508CCa1uIZM8cv8xqNGylB9w=">AAAB9HicbVDLSgMxFL1TX7W+qi7dBIvgqswUQRcuCi50JRXsA9qhZNJMG5pkxiRTKEO/w40LRdz6Me78GzPtLLT1QOBwzr3ckxPEnGnjut9OYW19Y3OruF3a2d3bPygfHrV0lChCmyTikeoEWFPOJG0aZjjtxIpiEXDaDsY3md+eUKVZJB/NNKa+wEPJQkawsZLfE9iMlEhv72f9Wr9ccavuHGiVeDmpQI5Gv/zVG0QkEVQawrHWXc+NjZ9iZRjhdFbqJZrGmIzxkHYtlVhQ7afz0DN0ZpUBCiNlnzRorv7eSLHQeioCO5mF1MteJv7ndRMTXvkpk3FiqCSLQ2HCkYlQ1gAaMEWJ4VNLMFHMZkVkhBUmxvZUsiV4y19eJa1a1XOr3sNFpX6d11GEEziFc/DgEupwBw1oAoEneIZXeHMmzovz7nwsRgtOvnMMf+B8/gCiYZH5</latexit>
...
<latexit sha1_base64="Gj7yv98SlyD93Ghofp+NnyXvd2c=">AAAB7HicbVBNS8NAFHypX7V+VT16WSyCp5KIUA8eCl48VjBtoQ1ls920SzebsPsilNDf4MWDIl79Qd78N27bHLR1YGGYecO+N2EqhUHX/XZKG5tb2zvl3cre/sHhUfX4pG2STDPus0QmuhtSw6VQ3EeBkndTzWkcSt4JJ3dzv/PEtRGJesRpyoOYjpSIBKNoJb8/TNAMqjW37i5A1olXkBoUaA2qXzbHspgrZJIa0/PcFIOcahRM8lmlnxmeUjahI96zVNGYmyBfLDsjF1YZkijR9ikkC/V3IqexMdM4tJMxxbFZ9ebif14vw+gmyIVKM+SKLT+KMkkwIfPLyVBozlBOLaFMC7srYWOqKUPbT8WW4K2evE7aV3XPrXsP17XmbVFHGc7gHC7BgwY04R5a4AMDAc/wCm+Ocl6cd+djOVpyiswp/IHz+QPvmo68</latexit>
GNM
<latexit sha1_base64="1uUQuLnXmq2FrQq5fvsHBbzm7v8=">AAAB9HicbVDLSgMxFL1TX7W+qi7dBIvgqsyIoAsXBRe6USrYB7RDyaSZNjTJjEmmUIZ+hxsXirj1Y9z5N2baWWjrgcDhnHu5JyeIOdPGdb+dwsrq2vpGcbO0tb2zu1feP2jqKFGENkjEI9UOsKacSdowzHDajhXFIuC0FYyuM781pkqzSD6aSUx9gQeShYxgYyW/K7AZKpHe3E97d71yxa26M6Bl4uWkAjnqvfJXtx+RRFBpCMdadzw3Nn6KlWGE02mpm2gaYzLCA9qxVGJBtZ/OQk/RiVX6KIyUfdKgmfp7I8VC64kI7GQWUi96mfif10lMeOmnTMaJoZLMD4UJRyZCWQOozxQlhk8swUQxmxWRIVaYGNtTyZbgLX55mTTPqp5b9R7OK7WrvI4iHMExnIIHF1CDW6hDAwg8wTO8wpszdl6cd+djPlpw8p1D+APn8wfLTZIU</latexit>
GM
<latexit sha1_base64="vjTCpRgsPEJfhljVzwQb7AFhV5c=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoNgFe5E0MIiYKGNENF8QHKEvc1csmRv79jdE8KRn2BjoYitv8jOf+MmuUITHww83pthZl6QCK6N6347hZXVtfWN4mZpa3tnd6+8f9DUcaoYNlgsYtUOqEbBJTYMNwLbiUIaBQJbweh66reeUGkey0czTtCP6EDykDNqrPRw07vrlStu1Z2BLBMvJxXIUe+Vv7r9mKURSsME1brjuYnxM6oMZwInpW6qMaFsRAfYsVTSCLWfzU6dkBOr9EkYK1vSkJn6eyKjkdbjKLCdETVDvehNxf+8TmrCSz/jMkkNSjZfFKaCmJhM/yZ9rpAZMbaEMsXtrYQNqaLM2HRKNgRv8eVl0jyrem7Vuz+v1K7yOIpwBMdwCh5cQA1uoQ4NYDCAZ3iFN0c4L8678zFvLTj5zCH8gfP5A+mdjYU=</latexit>
GNcore Gthid1 GNcore Gthid

<latexit sha1_base64="sfcetjjriA53KVhP8LRkSGs9KNA=">AAAB+3icbVDLSsNAFJ34rPUV69LNYBFclUQEXbgouNCVVLAPaEOYTCft0HmEmYlYQn7FjQtF3Poj7vwbJ20W2npg4HDOvdwzJ0oY1cbzvp2V1bX1jc3KVnV7Z3dv3z2odbRMFSZtLJlUvQhpwqggbUMNI71EEcQjRrrR5Lrwu49EaSrFg5kmJOBoJGhMMTJWCt3agCMzVjy7ucvDDEtF8tCtew1vBrhM/JLUQYlW6H4NhhKnnAiDGdK673uJCTKkDMWM5NVBqkmC8ASNSN9SgTjRQTbLnsMTqwxhLJV9wsCZ+nsjQ1zrKY/sZJFUL3qF+J/XT018GWRUJKkhAs8PxSmDRsKiCDikimDDppYgrKjNCvEYKYSNratqS/AXv7xMOmcN32v49+f15lVZRwUcgWNwCnxwAZrgFrRAG2DwBJ7BK3hzcufFeXc+5qMrTrlzCP7A+fwBopiUyw==</latexit>
<latexit sha1_base64="TTYaLKTJF7FnWkmBKBRepTsjYHc=">AAAB/3icbVDLSsNAFJ34rPUVFdy4GSyCG0sigi5cFFzosoJ9QBvDZDJph85MwsxEKDELf8WNC0Xc+hvu/BsnbRbaemDgcM693DMnSBhV2nG+rYXFpeWV1cpadX1jc2vb3tltqziVmLRwzGLZDZAijArS0lQz0k0kQTxgpBOMrgq/80CkorG40+OEeBwNBI0oRtpIvr1/fZ/pEzf3sz5Heih5NqRhnvt2zak7E8B54pakBko0ffurH8Y45URozJBSPddJtJchqSlmJK/2U0UShEdoQHqGCsSJ8rJJ/hweGSWEUSzNExpO1N8bGeJKjXlgJouQatYrxP+8XqqjCy+jIkk1EXh6KEoZ1DEsyoAhlQRrNjYEYUlNVoiHSCKsTWVVU4I7++V50j6tu07dvT2rNS7LOirgAByCY+CCc9AAN6AJWgCDR/AMXsGb9WS9WO/Wx3R0wSp39sAfWJ8/b2WWWQ==</latexit> <latexit sha1_base64="Vr1kZJh7jApVU4cGmtPimK4PPic=">AAAB/XicbVDLSsNAFJ3UV62v+Ni5GSyCq5KIoAsXBRe6rGAf0MYwmUzaoTOTMDMRagj+ihsXirj1P9z5N07aLLT1wMDhnHu5Z06QMKq043xblaXlldW16nptY3Nre8fe3euoOJWYtHHMYtkLkCKMCtLWVDPSSyRBPGCkG4yvCr/7QKSisbjTk4R4HA0FjShG2ki+fXB9n+nczwYc6ZHk2YiGee7bdafhTAEXiVuSOijR8u2vQRjjlBOhMUNK9V0n0V6GpKaYkbw2SBVJEB6jIekbKhAnysum6XN4bJQQRrE0T2g4VX9vZIgrNeGBmSxCqnmvEP/z+qmOLryMiiTVRODZoShlUMewqAKGVBKs2cQQhCU1WSEeIYmwNoXVTAnu/JcXSee04ToN9/as3rws66iCQ3AEToALzkET3IAWaAMMHsEzeAVv1pP1Yr1bH7PRilXu7IM/sD5/AI1Vlec=</latexit>
⇥M
<latexit sha1_base64="xCEPSgjeJaAOppNxwTZXrwRukIg=">AAAB73icbVA9SwNBEJ2LXzF+RS1tFoNgFe5E0MIiYGMjRDAfkBxhb7OXLNnbu+zOCeHIn7CxUMTWv2Pnv3GTXKGJDwYe780wMy9IpDDout9OYW19Y3OruF3a2d3bPygfHjVNnGrGGyyWsW4H1HApFG+gQMnbieY0CiRvBaPbmd964tqIWD3iJOF+RAdKhIJRtFK7iyLihtz3yhW36s5BVomXkwrkqPfKX91+zNKIK2SSGtPx3AT9jGoUTPJpqZsanlA2ogPesVRRu8bP5vdOyZlV+iSMtS2FZK7+nshoZMwkCmxnRHFolr2Z+J/XSTG89jOhkhS5YotFYSoJxmT2POkLzRnKiSWUaWFvJWxINWVoIyrZELzll1dJ86LquVXv4bJSu8njKMIJnMI5eHAFNbiDOjSAgYRneIU3Z+y8OO/Ox6K14OQzx/AHzucPqJqPrw==</latexit>
⇥M
GNenc
<latexit sha1_base64="KZY5NxgXVEC/q8QikcRvAbEAxBs=">AAAB+nicbVDLSsNAFL3xWesr1aWbYBFclUQEXbgouNCVVLAPaEOYTKft0JlJmJkoJeZT3LhQxK1f4s6/cRKz0NYDA4dz7mXOPWHMqNKu+2UtLa+srq1XNqqbW9s7u3Ztr6OiRGLSxhGLZC9EijAqSFtTzUgvlgTxkJFuOL3M/e49kYpG4k7PYuJzNBZ0RDHSRgrs2oAjPZE8vbrJgpQInAV23W24BZxF4pWkDiVagf05GEY44URozJBSfc+NtZ8iqSlmJKsOEkVihKdoTPqGCsSJ8tMieuYcGWXojCJpntBOof7eSBFXasZDM5kHVfNeLv7n9RM9OvdTKuJE51cVH40S5ujIyXtwhlQSrNnMEIQlNVkdPEESYW3aqpoSvPmTF0nnpOG5De/2tN68KOuowAEcwjF4cAZNuIYWtAHDAzzBC7xaj9az9Wa9/4wuWeXOPvyB9fENyBqUTg==</latexit>
GNdec
<latexit sha1_base64="79QHnx4t/4kSfeuqQfRiz+zDMfI=">AAAB+nicbVDLSsNAFJ3UV62vVJduBovgqiQi6MJFwYWupIJ9QBvCZDJph85MwsxEKTGf4saFIm79Enf+jZM2C209MHA4517umRMkjCrtON9WZWV1bX2julnb2t7Z3bPr+10VpxKTDo5ZLPsBUoRRQTqaakb6iSSIB4z0gslV4fceiFQ0Fvd6mhCPo5GgEcVIG8m360OO9Fjy7Po297OQ4Ny3G07TmQEuE7ckDVCi7dtfwzDGKSdCY4aUGrhOor0MSU0xI3ltmCqSIDxBIzIwVCBOlJfNoufw2CghjGJpntBwpv7eyBBXasoDM1kEVYteIf7nDVIdXXgZFUmqicDzQ1HKoI5h0QMMqSRYs6khCEtqskI8RhJhbdqqmRLcxS8vk+5p03Wa7t1Zo3VZ1lEFh+AInAAXnIMWuAFt0AEYPIJn8ArerCfrxXq3PuajFavcOQB/YH3+ALjdlEQ=</latexit>
GNenc
<latexit sha1_base64="KZY5NxgXVEC/q8QikcRvAbEAxBs=">AAAB+nicbVDLSsNAFL3xWesr1aWbYBFclUQEXbgouNCVVLAPaEOYTKft0JlJmJkoJeZT3LhQxK1f4s6/cRKz0NYDA4dz7mXOPWHMqNKu+2UtLa+srq1XNqqbW9s7u3Ztr6OiRGLSxhGLZC9EijAqSFtTzUgvlgTxkJFuOL3M/e49kYpG4k7PYuJzNBZ0RDHSRgrs2oAjPZE8vbrJgpQInAV23W24BZxF4pWkDiVagf05GEY44URozJBSfc+NtZ8iqSlmJKsOEkVihKdoTPqGCsSJ8tMieuYcGWXojCJpntBOof7eSBFXasZDM5kHVfNeLv7n9RM9OvdTKuJE51cVH40S5ujIyXtwhlQSrNnMEIQlNVkdPEESYW3aqpoSvPmTF0nnpOG5De/2tN68KOuowAEcwjF4cAZNuIYWtAHDAzzBC7xaj9az9Wa9/4wuWeXOPvyB9fENyBqUTg==</latexit>
GNdec
<latexit sha1_base64="79QHnx4t/4kSfeuqQfRiz+zDMfI=">AAAB+nicbVDLSsNAFJ3UV62vVJduBovgqiQi6MJFwYWupIJ9QBvCZDJph85MwsxEKTGf4saFIm79Enf+jZM2C209MHA4517umRMkjCrtON9WZWV1bX2julnb2t7Z3bPr+10VpxKTDo5ZLPsBUoRRQTqaakb6iSSIB4z0gslV4fceiFQ0Fvd6mhCPo5GgEcVIG8m360OO9Fjy7Po297OQ4Ny3G07TmQEuE7ckDVCi7dtfwzDGKSdCY4aUGrhOor0MSU0xI3ltmCqSIDxBIzIwVCBOlJfNoufw2CghjGJpntBwpv7eyBBXasoDM1kEVYteIf7nDVIdXXgZFUmqicDzQ1HKoI5h0QMMqSRYs6khCEtqskI8RhJhbdqqmRLcxS8vk+5p03Wa7t1Zo3VZ1lEFh+AInAAXnIMWuAFt0AEYPIJn8ArerCfrxXq3PuajFavcOQB/YH3+ALjdlEQ=</latexit>
G0
<latexit sha1_base64="vj48jMMQe2f55rU6zb6RZp+K9y4=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiIVFwELLiOYDkiPsbfaSJXt7x+6cEI78BBsLRWz9RXb+GzfJFZr4YODx3gwz84JECoOu++0U1tY3NreK26Wd3b39g/LhUcvEqWa8yWIZ605ADZdC8SYKlLyTaE6jQPJ2ML6Z+e0nro2I1SNOEu5HdKhEKBhFKz3c9t1+ueJW3TnIKvFyUoEcjX75qzeIWRpxhUxSY7qem6CfUY2CST4t9VLDE8rGdMi7lioaceNn81On5MwqAxLG2pZCMld/T2Q0MmYSBbYzojgyy95M/M/rphhe+ZlQSYpcscWiMJUEYzL7mwyE5gzlxBLKtLC3EjaimjK06ZRsCN7yy6ukdVH13Kp3f1mpX+dxFOEETuEcPKhBHe6gAU1gMIRneIU3RzovzrvzsWgtOPnMMfyB8/kDvamNaA==</latexit>
GNcore
GM
<latexit sha1_base64="vjTCpRgsPEJfhljVzwQb7AFhV5c=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoNgFe5E0MIiYKGNENF8QHKEvc1csmRv79jdE8KRn2BjoYitv8jOf+MmuUITHww83pthZl6QCK6N6347hZXVtfWN4mZpa3tnd6+8f9DUcaoYNlgsYtUOqEbBJTYMNwLbiUIaBQJbweh66reeUGkey0czTtCP6EDykDNqrPRw07vrlStu1Z2BLBMvJxXIUe+Vv7r9mKURSsME1brjuYnxM6oMZwInpW6qMaFsRAfYsVTSCLWfzU6dkBOr9EkYK1vSkJn6eyKjkdbjKLCdETVDvehNxf+8TmrCSz/jMkkNSjZfFKaCmJhM/yZ9rpAZMbaEMsXtrYQNqaLM2HRKNgRv8eVl0jyrem7Vuz+v1K7yOIpwBMdwCh5cQA1uoQ4NYDCAZ3iFN0c4L8678zFvLTj5zCH8gfP5A+mdjYU=</latexit>
⇥M
Ginp Gout Gtinp Gtout
<latexit sha1_base64="TKn8tu9S9/KYM9INqkcELhgYcuA=">AAAB+XicbVDLSsNAFJ3UV62vqEs3wSK4KokIunBRcKHLCvYBbQiT6aQdOo8wc1MooX/ixoUibv0Td/6NkzYLrR4YOJxzL/fMiVPODPj+l1NZW9/Y3Kpu13Z29/YP3MOjjlGZJrRNFFe6F2NDOZO0DQw47aWaYhFz2o0nt4XfnVJtmJKPMEtpKPBIsoQRDFaKXPcuygcCw1iLXGUwn0du3W/4C3h/SVCSOirRitzPwVCRTFAJhGNj+oGfQphjDYxwOq8NMkNTTCZ4RPuWSiyoCfNF8rl3ZpWhlyhtnwRvof7cyLEwZiZiO1mENKteIf7n9TNIrsOcyTQDKsnyUJJxD5RX1OANmaYE+MwSTDSzWT0yxhoTsGXVbAnB6pf/ks5FI/AbwcNlvXlT1lFFJ+gUnaMAXaEmukct1EYETdETekGvTu48O2/O+3K04pQ7x+gXnI9vWWaUGA==</latexit>
<latexit sha1_base64="Fm4Iz64LvWn082kf2+Dq6sd1g9I=">AAAB+3icbVDLSsNAFJ34rPUV69LNYBFclUQEXbgouNBlBfuANobJdNIOnWTCzI1YQn7FjQtF3Poj7vwbJ20W2npg4HDOvdwzJ0gE1+A439bK6tr6xmZlq7q9s7u3bx/UOlqmirI2lUKqXkA0EzxmbeAgWC9RjESBYN1gcl343UemNJfxPUwT5kVkFPOQUwJG8u3azQP42SAiMFZRJlPIc9+uOw1nBrxM3JLUUYmWb38NhpKmEYuBCqJ133US8DKigFPB8uog1SwhdEJGrG9oTCKmvWyWPccnRhniUCrzYsAz9fdGRiKtp1FgJouQetErxP+8fgrhpZfxOEmBxXR+KEwFBomLIvCQK0ZBTA0hVHGTFdMxUYSCqatqSnAXv7xMOmcN12m4d+f15lVZRwUdoWN0ilx0gZroFrVQG1H0hJ7RK3qzcuvFerc+5qMrVrlziP7A+vwB7zuU/g==</latexit>
<latexit sha1_base64="mvYjY6mgtt6w2Efm7YMP3XaOQ7I=">AAAB+XicbVDLSgMxFL1TX7W+Rl26CRbBVZkRQRcuCi50WcE+oB2GTJppQ5PMkGQKZeifuHGhiFv/xJ1/Y6adhVYPBA7n3EvOPVHKmTae9+VU1tY3Nreq27Wd3b39A/fwqKOTTBHaJglPVC/CmnImadsww2kvVRSLiNNuNLkt/O6UKs0S+WhmKQ0EHkkWM4KNlULXvQvzgcBmrETOZDqfh27da3gLoL/EL0kdSrRC93MwTEgmqDSEY637vpeaIMfKMMLpvDbINE0xmeAR7VsqsaA6yBfJ5+jMKkMUJ8o+adBC/bmRY6H1TER2sgipV71C/M/rZya+DoqLMkMlWX4UZxyZBBU1oCFTlBg+swQTxWxWRMZYYWJsWTVbgr968l/SuWj4XsN/uKw3b8o6qnACp3AOPlxBE+6hBW0gMIUneIFXJ3eenTfnfTlaccqdY/gF5+MbP22UBw==</latexit>
<latexit sha1_base64="KNSYXpe+JERsrbP/+QxnbJ2F1T0=">AAAB+3icbVDLSsNAFL3xWesr1qWbYBFclUQEXbgouNBlBfuANobJdNIOnZmE
(a) Composition of GN blocks (b) Encode-process-decode (c) Recurrent GN arch tecture
Figure 6: (a) An example composing multiple GN blocks in sequence to form a GN “core”. Here,
the GN b ocks can use shared we ghts or they cou d be ndependent (b) The encode-process-decode
architecture, which is a common choice for composing GN blocks (see Section 4.3). Here, a GN
encodes an input graph, which is then processed by a GN core. The output of the core is decoded
by a third GN block into an output graph, whose nodes, edges, and/or global attributes would be
used for task-specific purposes. (c) The encode-process-decode architecture applied in a sequential
sett ng n wh ch the core s a so unro ed over t me (potent a y us ng a GRU or LSTM arch tecture)
n add t on to be ng repeated w th n each t me step Here merged nes nd cate concatenat on and
split lines indicate copying.
4.3 Composable multi-block architectures

A key design principle of graph networks is constructing complex architectures by composing GN
blocks. We defined a GN block as always taking a graph comprised of edge, node, and global
elements as input, and returning a graph with the same constituent elements as output (simply
passing through the input elements to the output when those elements are not explicitly updated).
Th s graph-to-graph nput/output nterface ensures that the output of one GN b ock can be passed
as nput to another even f the r nterna configurat ons are d fferent s m ar to the tensor-to-tensor
interface of the standard deep learning toolkit. In the most basic form, two GN blocks, GN1 and
GN2 , can be composed as GN1 ◦ GN2 by passing the output of the first as input to the second:
G = GN2 (GN1 (G)).
Arbitrary numbers of GN blocks can be composed, as show in Figure 6a. The blocks can
be unshared (different functions and/or parameters, analogous to layers of a CNN), GN1 =
GN2 = · · · = GNM , or shared (reused functions and parameters, analogous to an unrolled RNN),
GN1 = GN2 = · · · = GNM . The white box around the GNcore in Figure 6a represents M repeated
internal processing sub-steps, with either shared or unshared GN blocks. Shared configurations
are analogous to message-passing (Gilmer et al., 2017), where the same local update procedure is
app ed terat ve y to propagate nformat on across the structure (F gure 7) If we exc ude the g oba
u (which aggregates information from across the nodes and edges), the information that a node
has access to after m steps of propagation is determined by the set of nodes and edges that are at
most m hops away. This can be interpreted as breaking down a complex computation into smaller
elementary steps. The steps can also be used to capture sequentiality in time. In our ball-spring
examp e f each propagat on step pred cts the phys ca dynam cs over one t me step of durat on ∆t
then the M propagation steps result in a total simulation time of, M ·∆t.
A common arch tecture des gn s what we ca the encode-process-decode configurat on (Hamr ck
et a (2018) a so see F gure 6ba) an nput graph G np s transformed nto a atent representat on
G0 , by an encoder, GNenc ; a shared core block, GNcore , is applied M times to return GM ; and
finally an output graph, Gout , is decoded by GNdec . For example, in our running example, the
encoder m ght compute the n t a forces and nteract on energ es between the ba s the core m ght
19
<latexit sha1_base64="ic6WezPV7TWar890N1QnpVOPjNA=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoOQU9gVQS9CwIvHiOYByRJmJ7PJkJnZZaZXCCGf4MWDIl79Im/+jZNkDxotaCiquunuilIpLPr+l1dYW9/Y3Cpul3Z29/YPyodHLZtkhvEmS2RiOhG1XArNmyhQ8k5qOFWR5O1ofDP324/cWJHoB5ykPFR0qEUsGEUn3atrv1+u+DV/AfKXBDmpQI5Gv/zZGyQsU1wjk9TabuCnGE6pQcEkn5V6meUpZWM65F1HNVXchtPFqTNy5pQBiRPjSiNZqD8nplRZO1GR61QUR3bVm4v/ed0M46twKnSaIddsuSjOJMGEzP8mA2E4QzlxhDIj3K2EjaihDF06JRdCsPryX9I6rwV+Lbi7qNSreRxFOIFTqEIAl1CHW2hAExgM4Qle4NWT3rP35r0vWwtePnMMv+B9fAO9341Y</latexit>
m=0 <latexit sha1_base64="Zv5ybfKo/H8QvXQaOWqQVyOFtg4=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoOQU9gVQS9CwIvHiOYByRJmJ73JkJnZZWZWCCGf4MWDIl79Im/+jZNkDxotaCiquunuilLBjfX9L6+wtr6xuVXcLu3s7u0flA+PWibJNMMmS0SiOxE1KLjCpuVWYCfVSGUksB2Nb+Z++xG14Yl6sJMUQ0mHisecUeuke3kd9MsVv+YvQP6SICcVyNHolz97g4RlEpVlghrTDfzUhlOqLWcCZ6VeZjClbEyH2HVUUYkmnC5OnZEzpwxInGhXypKF+nNiSqUxExm5TkntyKx6c/E/r5vZ+CqccpVmFhVbLoozQWxC5n+TAdfIrJg4Qpnm7lbCRlRTZl06JRdCsPryX9I6rwV+Lbi7qNSreRxFOIFTqEIAl1CHW2hAExgM4Qle4NUT3rP35r0vWwtePnMMv+B9fAO/Y41Z</latexit>
sha1_base64="41wuc+jmacSzA8ipzMzk3JQ30Cs=">AAAB33icbZDNSgMxFIXv1L9aq1a3boJFcFUybnQjCG5cVrS10A4lk95pQ5PMkGSEMvQR3LhQxLdy59uY/iy09UDg45yE3HviTArrKP0OShubW9s75d3KXnX/4LB2VG3bNDccWzyVqenEzKIUGltOOImdzCBTscSneHw7y5+e0ViR6kc3yTBSbKhFIjhz3npQ12G/VqcNOhdZh3AJdViq2a999QYpzxVqxyWzthvSzEUFM05widNKL7eYMT5mQ+x61EyhjYr5qFNy5p0BSVLjj3Zk7v5+UTBl7UTF/qZibmRXs5n5X9bNXXIVFUJnuUPNFx8luSQuJbO9yUAY5E5OPDBuhJ+V8BEzjDvfTsWXEK6uvA7ti0ZIG+E9hTKcwCmcQwiXcAN30IQWcBjCC7zBeyCD1+BjUVcpWPZ2DH8UfP4Arc6MHg==</latexit>
sha1_base64="c5O5JuZLnHr2YYnYcAr/a2hwAn0=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoOQKtzZaCMEbCwjmg9IjrC3mSRLdveO3T0hHPkJNhaK2PqL7Pw3bpIrNPHBwOO9GWbmRYngxvr+t1fY2Nza3inulvb2Dw6PyscnLROnmmGTxSLWnYgaFFxh03IrsJNopDIS2I4mt3O//YTa8Fg92mmCoaQjxYecUeukB3kT9MsVv+YvQNZJkJMK5Gj0y1+9QcxSicoyQY3pBn5iw4xqy5nAWamXGkwom9ARdh1VVKIJs8WpM3LhlAEZxtqVsmSh/p7IqDRmKiPXKakdm1VvLv7ndVM7vA4zrpLUomLLRcNUEBuT+d9kwDUyK6aOUKa5u5WwMdWUWZdOyYUQrL68TlqXtcCvBfd+pV7N4yjCGZxDFQK4gjrcQQOawGAEz/AKb57wXrx372PZWvDymVP4A+/zB74jjVU=</latexit>
m=1 <latexit sha1_base64="oVNcNejQAmVb9Nn6VIeeV7f50ls=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvQU9ktgl6EghePFe0HtEvJptk2NMkuSVYoS3+CFw+KePUXefPfmLZ70NYHA4/3ZpiZFyaCG+t536iwsbm1vVPcLe3tHxwelY9P2iZONWUtGotYd0NimOCKtSy3gnUTzYgMBeuEk9u533li2vBYPdppwgJJRopHnBLrpAd5Ux+UK17NWwCvEz8nFcjRHJS/+sOYppIpSwUxpud7iQ0yoi2ngs1K/dSwhNAJGbGeo4pIZoJsceoMXzhliKNYu1IWL9TfExmRxkxl6DolsWOz6s3F/7xeaqPrIOMqSS1TdLkoSgW2MZ7/jYdcM2rF1BFCNXe3YjommlDr0im5EPzVl9dJu17zvZp/f1lpVPM4inAG51AFH66gAXfQhBZQGMEzvMIbEugFvaOPZWsB5TOn8Afo8wfA541a</latexit>
m=2 <latexit sha1_base64="+jOsYdJ9SNJymWDVPaXJffNRSPM=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvQU9lVQS9CwYvHivYD2qVk02wbmmSXJCuUpT/BiwdFvPqLvPlvTNs9aOuDgcd7M8zMCxPBjfW8b1RYW9/Y3Cpul3Z29/YPyodHLROnmrImjUWsOyExTHDFmpZbwTqJZkSGgrXD8e3Mbz8xbXisHu0kYYEkQ8UjTol10oO8ueiXK17NmwOvEj8nFcjR6Je/eoOYppIpSwUxput7iQ0yoi2ngk1LvdSwhNAxGbKuo4pIZoJsfuoUnzllgKNYu1IWz9XfExmRxkxk6DolsSOz7M3E/7xuaqPrIOMqSS1TdLEoSgW2MZ79jQdcM2rFxBFCNXe3YjoimlDr0im5EPzll1dJ67zmezX//rJSr+ZxFOEETqEKPlxBHe6gAU2gMIRneIU3JNALekcfi9YCymeO4Q/Q5w/Ca41b</latexit>
m=3
<latexit sha1_base64="ic6WezPV7TWar890N1QnpVOPjNA=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoOQU9gVQS9CwIvHiOYByRJmJ7PJkJnZZaZXCCGf4MWDIl79Im/+jZNkDxotaCiquunuilIpLPr+l1dYW9/Y3Cpul3Z29/YPyodHLZtkhvEmS2RiOhG1XArNmyhQ8k5qOFWR5O1ofDP324/cWJHoB5ykPFR0qEUsGEUn3atrv1+u+DV/AfKXBDmpQI5Gv/zZGyQsU1wjk9TabuCnGE6pQcEkn5V6meUpZWM65F1HNVXchtPFqTNy5pQBiRPjSiNZqD8nplRZO1GR61QUR3bVm4v/ed0M46twKnSaIddsuSjOJMGEzP8mA2E4QzlxhDIj3K2EjaihDF06JRdCsPryX9I6rwV+Lbi7qNSreRxFOIFTqEIAl1CHW2hAExgM4Qle4NWT3rP35r0vWwtePnMMv+B9fAO9341Y</latexit>
m=0 <latexit sha1_base64="Zv5ybfKo/H8QvXQaOWqQVyOFtg4=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoOQU9gVQS9CwIvHiOYByRJmJ73JkJnZZWZWCCGf4MWDIl79Im/+jZNkDxotaCiquunuilLBjfX9L6+wtr6xuVXcLu3s7u0flA+PWibJNMMmS0SiOxE1KLjCpuVWYCfVSGUksB2Nb+Z++xG14Yl6sJMUQ0mHisecUeuke3kd9MsVv+YvQP6SICcVyNHolz97g4RlEpVlghrTDfzUhlOqLWcCZ6VeZjClbEyH2HVUUYkmnC5OnZEzpwxInGhXypKF+nNiSqUxExm5TkntyKx6c/E/r5vZ+CqccpVmFhVbLoozQWxC5n+TAdfIrJg4Qpnm7lbCRlRTZl06JRdCsPryX9I6rwV+Lbi7qNSreRxFOIFTqEIAl1CHW2hAExgM4Qle4NUT3rP35r0vWwtePnMMv+B9fAO/Y41Z</latexit>
sha1_base64="41wuc+jmacSzA8ipzMzk3JQ30Cs=">AAAB33icbZDNSgMxFIXv1L9aq1a3boJFcFUybnQjCG5cVrS10A4lk95pQ5PMkGSEMvQR3LhQxLdy59uY/iy09UDg45yE3HviTArrKP0OShubW9s75d3KXnX/4LB2VG3bNDccWzyVqenEzKIUGltOOImdzCBTscSneHw7y5+e0ViR6kc3yTBSbKhFIjhz3npQ12G/VqcNOhdZh3AJdViq2a999QYpzxVqxyWzthvSzEUFM05widNKL7eYMT5mQ+x61EyhjYr5qFNy5p0BSVLjj3Zk7v5+UTBl7UTF/qZibmRXs5n5X9bNXXIVFUJnuUPNFx8luSQuJbO9yUAY5E5OPDBuhJ+V8BEzjDvfTsWXEK6uvA7ti0ZIG+E9hTKcwCmcQwiXcAN30IQWcBjCC7zBeyCD1+BjUVcpWPZ2DH8UfP4Arc6MHg==</latexit>
sha1_base64="c5O5JuZLnHr2YYnYcAr/a2hwAn0=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoOQKtzZaCMEbCwjmg9IjrC3mSRLdveO3T0hHPkJNhaK2PqL7Pw3bpIrNPHBwOO9GWbmRYngxvr+t1fY2Nza3inulvb2Dw6PyscnLROnmmGTxSLWnYgaFFxh03IrsJNopDIS2I4mt3O//YTa8Fg92mmCoaQjxYecUeukB3kT9MsVv+YvQNZJkJMK5Gj0y1+9QcxSicoyQY3pBn5iw4xqy5nAWamXGkwom9ARdh1VVKIJs8WpM3LhlAEZxtqVsmSh/p7IqDRmKiPXKakdm1VvLv7ndVM7vA4zrpLUomLLRcNUEBuT+d9kwDUyK6aOUKa5u5WwMdWUWZdOyYUQrL68TlqXtcCvBfd+pV7N4yjCGZxDFQK4gjrcQQOawGAEz/AKb57wXrx372PZWvDymVP4A+/zB74jjVU=</latexit>
m=1 <latexit sha1_base64="oVNcNejQAmVb9Nn6VIeeV7f50ls=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvQU9ktgl6EghePFe0HtEvJptk2NMkuSVYoS3+CFw+KePUXefPfmLZ70NYHA4/3ZpiZFyaCG+t536iwsbm1vVPcLe3tHxwelY9P2iZONWUtGotYd0NimOCKtSy3gnUTzYgMBeuEk9u533li2vBYPdppwgJJRopHnBLrpAd5Ux+UK17NWwCvEz8nFcjRHJS/+sOYppIpSwUxpud7iQ0yoi2ngs1K/dSwhNAJGbGeo4pIZoJsceoMXzhliKNYu1IWL9TfExmRxkxl6DolsWOz6s3F/7xeaqPrIOMqSS1TdLkoSgW2MZ7/jYdcM2rF1BFCNXe3YjommlDr0im5EPzVl9dJu17zvZp/f1lpVPM4inAG51AFH66gAXfQhBZQGMEzvMIbEugFvaOPZWsB5TOn8Afo8wfA541a</latexit>
m=2 <latexit sha1_base64="+jOsYdJ9SNJymWDVPaXJffNRSPM=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvQU9lVQS9CwYvHivYD2qVk02wbmmSXJCuUpT/BiwdFvPqLvPlvTNs9aOuDgcd7M8zMCxPBjfW8b1RYW9/Y3Cpul3Z29/YPyodHLROnmrImjUWsOyExTHDFmpZbwTqJZkSGgrXD8e3Mbz8xbXisHu0kYYEkQ8UjTol10oO8ueiXK17NmwOvEj8nFcjR6Je/eoOYppIpSwUxput7iQ0yoi2ngk1LvdSwhNAxGbKuo4pIZoJsfuoUnzllgKNYu1IWz9XfExmRxkxk6DolsSOz7M3E/7xuaqPrIOMqSS1TdLEoSgW2MZ79jQdcM2rFxBFCNXe3YjoimlDr0im5EPzll1dJ67zmezX//rJSr+ZxFOEETqEKPlxBHe6gAU2gMIRneIU3JNALekcfi9YCymeO4Q/Q5w/Ca41b</latexit>
m=3
Figure 7: Example of message passing. Each row highlights the information that diffuses through
the graph starting from a particular node. In the top row, the node of interest is in the upper
right; in the bottom row, the node of interest is in the bottom right. Shaded nodes indicate how far
information from the original node can travel in m steps of message passing; bolded edges indicate
which edges that information has the potential to travel across. Note that during the full message
passing procedure, this propagation of information happens simultaneously for all nodes and edges
in the graph (not just the two shown here).
apply an elementary dynamics update, and the decoder might read out the final positions from the
updated graph state.
Similar to the encode-process-decode design, recurrent GN-based architectures can be built by
maintaining a hidden graph, Gthid , taking as input an observed graph, Gtinp , and returning an output
graph, Gtout , on each step (see Figure 6c). This type of architecture can be particularly useful for
predicting sequences of graphs, such as predicting the trajectory of a dynamical system over time
(e.g. Sanchez-Gonzalez et al., 2018). The encoded graph, output by GNenc , must have the same
structure as Gthid , and they can be easily combined by concatenating their corresponding ek , vi ,
and u vectors (where the upward arrow merges into the left-hand horizontal arrow in Figure 6c),
before being passed to GNcore . For the output, the Gthid is copied (where the right-hand horizontal
arrow splits into the downward arrow in Figure 6c) and decoded by GNdec . This design reuses GN
blocks in several ways: GNenc , GNdec , and GNcore are shared across each step, t; and within each
step, GNcore may perform multiple shared sub-steps.
Various other techniques for designing GN-based architectures can be useful. Graph skip
connections, for example, would concatenate a GN block’s input graph, Gm , with its output graph,
Gm+1 , before proceeding to further computations. Merging and smoothing input and hidden graph
information, as in Figure 6c, can use LSTM- or GRU-style gating schemes, instead of simple
concatenation (Li et al., 2016). Or distinct, recurrent GN blocks (e.g. Figure 4b) can be composed
before and/or after other GN blocks, to improve stability in the representations over multiple
propagation steps (Sanchez-Gonzalez et al., 2018).
4.4 Implementing graph networks in code

Similar to CNNs (see Figure 1), which are naturally parallelizable (e.g. on GPUs), GNs have a
natural parallel structure: since the φe and φv functions in Equation 1 are shared over the edges
and nodes, respectively, they can be computed in parallel. In practice, this means that with respect
20
Box 4: Graph Nets open-source software library: github.com/deepmind/graph nets
We have released an open-source library for building GNs in Tensorflow/Sonnet. It includes
demos of how to create, manipulate, and train GNs to reason about graph-structured data, on
a shortest path-finding task, a sorting task, and a physical prediction task. Each demo uses the
same GN architecture, which highlights the flexibility of the approach.
Shortest path demo: tinyurl.com/gn-shortest-path-demo

This demo creates random graphs, and trains a GN to label the nodes and edges on the shortest
path between any two nodes. Over a sequence of message-passing steps (as depicted by each
step’s plot), the model refines its prediction of the shortest path.
Sort demo: tinyurl.com/gn-sort-demo

This demo creates lists of random numbers, and trains a GN to sort the list. After a sequence
of message-passing steps, the model makes an accurate prediction of which elements (columns
in the figure) come next after each other (rows).
Physics demo: tinyurl.com/gn-physics-demo

This demo creates random mass-spring physical systems, and trains a GN to predict the state of
the system on the next timestep. The model’s next-step predictions can be fed back in as input
to create a rollout of a future trajectory. Each subplot below shows the true and predicted
mass-spring system states over 50 timesteps. This is similar to the model and experiments in
(Battaglia et al., 2016)’s “interaction networks”.
21
to φe and φv , the nodes and edges can be treated like the batch dimension in typical mini-batch
training regimes. Moreover, several graphs can be naturally batched together by treating them as
disjoint components of a larger graph. With some additional bookkeeping, this allows batching
together the computations made on several independent graphs.
Reusing φe and φv also improves GNs’ sample efficiency. Again, analogous to a convolutional
kernel, the number of samples which are used to optimize a GN’s φe and φv functions is the number
of edges and nodes, respectively, across all training graphs. For example, in the balls example
from Sec. 3.2, a scene with four balls which are all connected by springs will provide twelve (4 × 3)
examples of the contact interaction between them.
We have released an open-source software library for building GNs, which can be found here:
github.com/deepmind/graph nets. See Box 4 for an overview.
4.5 Summary
In this section, we have discussed the design principles behind graph networks: flexible representa-
tions, configurable within-block structure, and composable multi-block architectures. These three
design principles combine in our framework which is extremely flexible and applicable to a wide range
of domains ranging from perception, language, and symbolic reasoning. And, as we will see in the
remainder of this paper, the strong relational inductive bias possessed by graph networks supports
combinatorial generalization, thus making it a powerful tool both in terms of implementation and
theory.
5 Discussion
In this paper, we analyzed the extent to which relational inductive bias exists in deep learning
architectures like MLPs, CNNs, and RNNs, and concluded that while CNNs and RNNs do contain
relational inductive biases, they cannot naturally handle more structured representations such as
sets or graphs. We advocated for building stronger relational inductive biases into deep learning
architectures by highlighting an underused deep learning building block called a graph network,
which performs computations over graph-structured data. Our graph network framework unifies
existing approaches that also operate over graphs, and provides a straightforward interface for
assembling graph networks into complex, sophisticated architectures.
5.1 Combinatorial generalization in graph networks

The structure of GNs naturally supports combinatorial generalization because they do not perform
computations strictly at the system level, but also apply shared computations across the entities and
across the relations as well. This allows never-before-seen systems to be reasoned about, because
they are built from familiar components, in a way that reflects von Humboldt’s “infinite use of finite
means” (Humboldt, 1836; Chomsky, 1965).
A number of studies have explored GNs’ capacity for combinatorial generalization. Battaglia
et al. (2016) found that GNs trained to make one-step physical state predictions could simulate
thousands of future time steps, and also exhibit accurate zero-shot transfer to physical systems
with double, or half, the number of entities experienced during training. Sanchez-Gonzalez et al.
(2018) found similar results in more complex physical control settings, including that GNs trained as
forward models on simulated multi-joint agents could generalize to agents with new numbers of joints.
Hamrick et al. (2018) and Wang et al. (2018b) each found that GN-based decision-making policies
could transfer to novel numbers of entities as well. In combinatorial optimization problems, Bello
22
et al. (2016); Nowak et al. (2017); Dai et al. (2017); Kool and Welling (2018) showed that GNs could
generalize well to problems of much different sizes than they had been trained on. Similarly, Toyer
et al. (2017) showed generalization to different sizes of planning problems, and Hamilton et al. (2017)
showed generalization to producing useful node embeddings for previously unseen data. On boolean
SAT problems, Selsam et al. (2018) demonstrated generalization both to different problem sizes and
across problem distributions: their model retained good performance upon strongly modifying the
distribution of the input graphs and its typical local structure.
These striking examples of combinatorial generalization are not entirely surprising, given GNs’
entity- and relation-centric organization, but nonetheless provide important support for the view
that embracing explicit structure and flexible learning is a viable approach toward realizing better
sample efficiency and generalization in modern AI.
5.2 Limitations of graph networks

One limitation of GNs’ and MPNNs’ form of learned message-passing (Shervashidze et al., 2011)
is that it cannot be guaranteed to solve some classes of problems, such as discriminating between
certain non-isomorphic graphs. Kondor et al. (2018) suggested that covariance7 (Cohen and Welling,
2016; Kondor and Trivedi, 2018), rather than invariance to permutations of the nodes and edges
is preferable, and proposed “covariant compositional networks” which can preserve structural
information, and allow it to be ignored only if desired.
More generally, while graphs are a powerful way of representing structure information, they
have limits. For example, notions like recursion, control flow, and conditional iteration are not
straightforward to represent with graphs, and, minimally, require additional assumptions (e.g., in
interpreting abstract syntax trees). Programs and more “computer-like” processing can offer greater
representational and computational expressivity with respect to these notions, and some have argued
they are an important component of human cognition (Tenenbaum et al., 2011; Lake et al., 2015;
Goodman et al., 2015).
5.3 Open questions

Although we are excited about the potential impacts that graph networks can have, we caution that
these models are only one step forward. Realizing the full potential of graph networks will likely be
far more challenging than organizing their behavior under one framework, and indeed, there are a
number of unanswered questions regarding the best ways to use graph networks.
One pressing question is: where do the graphs come from that graph networks operate over?
One of the hallmarks of deep learning has been its ability to perform complex computations over
raw sensory data, such as images and text, yet it is unclear the best ways to convert sensory data
into more structured representations like graphs. One approach (which we have already discussed)
assumes a fully connected graph structure between spatial or linguistic entities, such as in the
literature on self-attention (Vaswani et al., 2017; Wang et al., 2018c). However, such representations
may not correspond exactly to the “true” entities (e.g., convolutional features do not directly
correspond to objects in a scene). Moreover, many underlying graph structures are much more
sparse than a fully connected graph, and it is an open question how to induce this sparsity. Several
lines of active research are exploring these issues (Watters et al., 2017; van Steenkiste et al., 2018;
Li et al., 2018; Kipf et al., 2018) but as of yet there is no single method which can reliably extract
discrete entities from sensory data. Developing such a method is an exciting challenge for future
7
Covariance means, roughly, that the activations vary in a predictable way as a function of the ordering of the
incoming edges.
23
research, and once solved will likely open the door for much more powerful and flexible reasoning
algorithms.
A related question is how to adaptively modify graph structures during the course of computation.
For example, if an object fractures into multiple pieces, a node representing that object also ought
to split into multiple nodes. Similarly, it might be useful to only represent edges between objects
that are in contact, thus requiring the ability to add or remove edges depending on context. The
question of how to support this type of adaptivity is also actively being researched, and in particular,
some of the methods used for identifying the underlying structure of a graph may be applicable (e.g.
Li et al., 2018; Kipf et al., 2018).
Human cognition makes the strong assumption that the world is composed of objects and
relations (Spelke and Kinzler, 2007), and because GNs make a similar assumption, their behavior
tends to be more interpretable. The entities and relations that GNs operate over often correspond
to things that humans understand (such as physical objects), thus supporting more interpretable
analysis and visualization (e.g., as in Selsam et al., 2018). An interesting direction for future work
is to further explore the interpretability of the behavior of graph networks.
5.4 Integrative approaches for learning and structure

While our focus here has been on graphs, one takeaway from this paper is less about graphs
themselves and more about the approach of blending powerful deep learning approaches with
structured representations. We are excited by related approaches which have explored this idea for
other types of structured representations and computations, such as linguistic trees (Socher et al.,
2011a,b, 2012, 2013; Tai et al., 2015; Andreas et al., 2016), partial tree traversals in a state-action
graph (Guez et al., 2018; Farquhar et al., 2018), hierarchical action policies (Andreas et al., 2017),
multi-agent communication channels (Foerster et al., 2016), “capsules” (Sabour et al., 2017), and
programs (Parisotto et al., 2017). Other methods have attempted to capture different types of
structure by mimicking key hardware and software components in computers and how they transfer
information between each other, such as persistent slotted storage, registers, memory I/O controllers,
stacks, and queues (e.g. Dyer et al., 2015; Grefenstette et al., 2015; Joulin and Mikolov, 2015;
Sukhbaatar et al., 2015; Kurach et al., 2016; Graves et al., 2016).
5.5 Conclusion
Recent advances in AI, propelled by deep learning, have been transformative across many important
domains. Despite this, a vast gap between human and machine intelligence remains, especially with
respect to efficient, generalizable learning. We argue for making combinatorial generalization a top
priority for AI, and advocate for embracing integrative approaches which draw on ideas from human
cognition, traditional computer science, standard engineering practice, and modern deep learning.
Here we explored flexible learning-based approaches which implement strong relational inductive
biases to capitalize on explicitly structured representations and computations, and presented a
framework called graph networks, which generalize and extend various recent approaches for neural
networks applied to graphs. Graph networks are designed to promote building complex architectures
using customizable graph-to-graph building blocks, and their relational inductive biases promote
combinatorial generalization and improved sample efficiency over other standard machine learning
building blocks.
Despite their benefits and potential, however, learnable models which operate on graphs are
only a stepping stone on the path toward human-like intelligence. We are optimistic about a
number of other relevant, and perhaps underappreciated, research directions, including marrying
24
learning-based approaches with programs (Ritchie et al., 2016; Andreas et al., 2016; Gaunt et al.,
2016; Evans and Grefenstette, 2018; Evans et al., 2018), developing model-based approaches with an
emphasis on abstraction (Kansky et al., 2017; Konidaris et al., 2018; Zhang et al., 2018; Hay et al.,
2018), investing more heavily in meta-learning (Wang et al., 2016, 2018a; Finn et al., 2017), and
exploring multi-agent learning and interaction as a key catalyst for advanced intelligence (Nowak,
2006; Ohtsuki et al., 2006). These directions each involve rich notions of entities, relations, and
combinatorial generalization, and can potentially benefit, and benefit from, greater interaction with
approaches for learning relational reasoning over explicitly structured representations.
Acknowledgements
We thank Tobias Pfaff, Danilo Rezende, Nando de Freitas, Murray Shanahan, Thore Graepel, John
Jumper, Demis Hassabis, and the broader DeepMind and Google communities for valuable feedback
and support.
References
Allamanis, M., Brockschmidt, M., and Khademi, M. (2018). Learning to represent programs with
graphs. In Proceedings of the International Conference on Learning Representations (ICLR).
Allamanis, M., Chanthirasegaran, P., Kohli, P., and Sutton, C. (2017). Learning continuous
semantic representations of symbolic expressions. In Proceedings of the International Conference
on Machine Learning (ICML).
Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89(4):369.
Andreas, J., Klein, D., and Levine, S. (2017). Modular multitask reinforcement learning with policy
sketches. In Proceedings of the International Conference on Machine Learning (ICML).
Andreas, J., Rohrbach, M., Darrell, T., and Klein, D. (2016). Neural module networks. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 39–48.
Ba, J. L., Kiros, J. R., and Hinton, G. E. (2016). Layer normalization. arXiv preprint
arXiv:1607.06450.
Bahdanau, D., Cho, K., and Bengio, Y. (2015). Neural machine translation by jointly learning to
align and translate. In Proceedings of the International Conference on Learning Representations
(ICLR).
Battaglia, P., Pascanu, R., Lai, M., Rezende, D. J., et al. (2016). Interaction networks for learning
about objects, relations and physics. In Advances in Neural Information Processing Systems,
pages 4502–4510.
Battaglia, P. W., Hamrick, J. B., and Tenenbaum, J. B. (2013). Simulation as an engine of physical
scene understanding. Proceedings of the National Academy of Sciences, 110(45):18327–18332.
Bello, I., Pham, H., Le, Q. V., Norouzi, M., and Bengio, S. (2016). Neural combinatorial optimization
with reinforcement learning. arXiv preprint arXiv:1611.09940.
Bobrow, D. G. and Hinton, G. E., editors (1990). Artificial Intelligence, volume 46. Elsevier Science
Publishers Ltd., Essex, UK. Special Issue 1-2: On Connectionist Symbol Processing.
25
Bojchevski, A., Shchur, O., Zügner, D., and Günnemann, S. (2018). Netgan: Generating graphs via
random walks. arXiv preprint arXiv:1803.00816.
Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., and Yakhnenko, O. (2013). Translating
embeddings for modeling multi-relational data. In Advances in Neural Information Processing
Systems, pages 2787–2795.
Botvinick, M. M. (2008). Hierarchical models of behavior and prefrontal function. Trends in

Cognitive Sciences, 12(5):201–208.
Bronstein, M. M., Bruna, J., LeCun, Y., Szlam, A., and Vandergheynst, P. (2017). Geometric deep
learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42.
Bruna, J., Zaremba, W., Szlam, A., and LeCun, Y. (2014). Spectral networks and locally connected
networks on graphs. In Proceedings of the International Conference on Learning Representations
(ICLR).
Chang, M. B., Ullman, T., Torralba, A., and Tenenbaum, J. B. (2017). A compositional object-
based approach to learning physical dynamics. In Proceedings of the International Conference on
Learning Representations (ICLR).
Chen, X., Li, L., Fei-Fei, L., and Gupta, A. (2018a). Iterative visual reasoning beyond convolutions.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
Chen, X., Liu, C., and Song, D. (2018b). Tree-to-tree neural networks for program translation. In
Workshops of the International Conference on Learning Representations (ICLR).
Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., and Bengio,
Y. (2014). Learning phrase representations using rnn encoder-decoder for statistical machine
translation. In Proceeding of the Conference on Empirical Methods in Natural Language Processing
(EMNLP).
Chomsky, N. (1957). Syntactic Structures. Mouton & Co.
Chomsky, N. (1965). Aspects of the Theory of Syntax. MIT Press.
Cohen, T. and Welling, M. (2016). Group equivariant convolutional networks. In International

Conference on Machine Learning, pages 2990–2999.
Craik, K. J. W. (1943). The Nature of Explanation. Cambridge University Press.
Cui, Z., Henrickson, K., Ke, R., and Wang, Y. (2018). High-order graph convolutional recurrent
neural network: A deep learning framework for network-scale traffic learning and forecasting.
arXiv preprint arXiv:1802.07007.
Dai, H., Dai, B., and Song, L. (2016). Discriminative embeddings of latent variable models for
structured data. In Proceedings of the International Conference on Machine Learning (ICML).
Dai, H., Khalil, E. B., Zhang, Y., Dilkina, B., and Song, L. (2017). Learning combinatorial
optimization algorithms over graphs. In Advances in Neural Information Processing Systems.
De Cao, N. and Kipf, T. (2018). MolGAN: An implicit generative model for small molecular graphs.
26
Defferrard, M., Bresson, X., and Vandergheynst, P. (2016). Convolutional neural networks on graphs
with fast localized spectral filtering. In Advances in Neural Information Processing Systems, pages
3844–3852.
Denil, M., Colmenarejo, S. G., Cabi, S., Saxton, D., and de Freitas, N. (2017). Programmable
agents. arXiv preprint arXiv:1706.06383.
Devlin, J., Uesato, J., Singh, R., and Kohli, P. (2017). Semantic code repair using neuro-symbolic
transformation networks. arXiv preprint arXiv:1710.11054.
Duvenaud, D. K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., and
Adams, R. P. (2015). Convolutional networks on graphs for learning molecular fingerprints. In
Advances in Neural Information Processing Systems, pages 2224–2232.
Dyer, C., Ballesteros, M., Ling, W., Matthews, A., and Smith, N. A. (2015). Transition-based
dependency parsing with stack long short-term memory. In Proceedings of the Annual Meeting of
the Association for Computational Linguistics (ACL).
Džeroski, S., De Raedt, L., and Driessens, K. (2001). Relational reinforcement learning. Machine
Learning, 43(1-2):7–52.
Edwards, H. and Storkey, A. (2016). Towards a neural statistician. arXiv preprint arXiv:1606.02185.
Eliasmith, C. (2013). How to build a brain: A neural architecture for biological cognition. Oxford
University Press.
Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2):179–211.
Elman, J. L. (1991). Distributed representations, simple recurrent networks, and grammatical

structure. Machine Learning, 7(2-3):195–225.
Eslami, S. A., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Hinton, G. E., et al. (2016). Attend,
infer, repeat: Fast scene understanding with generative models. In Advances in Neural Information
Processing Systems, pages 3225–3233.
Evans, R. and Grefenstette, E. (2018). Learning explanatory rules from noisy data. Journal of
Artificial Intelligence Research, 61:1–64.
Evans, R., Saxton, D., Amos, D., Kohli, P., and Grefenstette, E. (2018). Can neural networks
understand logical entailment? In Proceedings of the International Conference on Learning
Representations (ICLR).
Farquhar, G., Rocktäschel, T., Igl, M., and Whiteson, S. (2018). TreeQN and ATreeC: Differentiable
tree planning for deep reinforcement learning. In Proceedings of the International Conference on
Finn, C., Abbeel, P., and Levine, S. (2017). Model-agnostic meta-learning for fast adaptation of
deep networks. arXiv preprint arXiv:1703.03400.
Fodor, J. A. (1975). The Language of Thought. Harvard University Press.
Fodor, J. A. and Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical

analysis. Cognition, 28(1-2):3–71.
27
Foerster, J., Assael, I. A., de Freitas, N., and Whiteson, S. (2016). Learning to communicate with
deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems,
pages 2137–2145.
Fukushima, K. (1980). Neocognitron: A self-organizing neural network model for a mechanism of
pattern recognition unaffected by shift in position. Biological Cybernetics, 36:193–202.
Garcia, V. and Bruna, J. (2018). Few-shot learning with graph neural networks. In Proceedings of
the International Conference on Learning Representations (ICLR).
Garcı́a-Durán, A. and Niepert, M. (2017). Learning graph representations with embedding propaga-
tion. arXiv preprint arXiv:1710.03059.
Garnelo, M., Arulkumaran, K., and Shanahan, M. (2016). Towards deep symbolic reinforcement
learning. arXiv preprint arXiv:1609.05518.
Gaunt, A. L., Brockschmidt, M., Kushman, N., and Tarlow, D. (2016). Differentiable programs
with neural libraries. arXiv preprint arXiv:1611.02109.
Geman, S., Bienenstock, E., and Doursat, R. (1992). Neural networks and the bias/variance dilemma.
Neural Computation, 4(1):1–58.
Gentner, D. and Markman, A. B. (1997). Structure mapping in analogy and similarity. American
Psychologist, 52(1):45.
Getoor, L. and Taskar, B. (2007). Introduction to Statistical Relational Learning. MIT press.
Ghahramani, Z. (2015). Probabilistic machine learning and artificial intelligence. Nature,
521(7553):452.
Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. (2017). Neural message
passing for quantum chemistry. arXiv preprint arXiv:1704.01212.
Goodfellow, I., Bengio, Y., Courville, A., and Bengio, Y. (2016). Deep Learning. MIT Press.
Goodman, N. (1955). The new riddle of induction. In Fact, Fiction, and Forecast, pages 59–83.
Harvard University Press.
Goodman, N., Mansinghka, V., Roy, D. M., Bonawitz, K., and Tenenbaum, J. B. (2012). Church: a
language for generative models. arXiv preprint arXiv:1206.3255.
Goodman, N. D., Tenenbaum, J. B., and Gerstenberg, T. (2015). Concepts in a probabilistic
language of thought. In Margolis, E. and Laurence, S., editors, The Conceptual Mind: New
Directions in the Study of Concepts. MIT Press.
Goodwin, G. P. and Johnson-Laird, P. (2005). Reasoning about relations. Psychological Review,
112(2):468.
Gori, M., Monfardini, G., and Scarselli, F. (2005). A new model for learning in graph domains. In
Proceedings of the International Joint Conference on Neural Networks (IJCNN), volume 2, pages
729–734. IEEE.
Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A., Colmenarejo,
S. G., Grefenstette, E., Ramalho, T., Agapiou, J., et al. (2016). Hybrid computing using a neural
network with dynamic external memory. Nature, 538(7626):471.
28
Grefenstette, E., Hermann, K. M., Suleyman, M., and Blunsom, P. (2015). Learning to transduce
with unbounded memory. In Advances in Neural Information Processing Systems, pages 1828–1836.
Griffiths, T. L., Chater, N., Kemp, C., Perfors, A., and Tenenbaum, J. B. (2010). Probabilistic
models of cognition: Exploring representations and inductive biases. Trends in Cognitive Sciences,
14(8):357–364.
Grover, A. and Leskovec, J. (2016). node2vec: Scalable feature learning for networks. In Proceedings
of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining,
pages 855–864. ACM.
Guez, A., Weber, T., Antonoglou, I., Simonyan, K., Vinyals, O., Wierstra, D., Munos, R., and
Silver, D. (2018). Learning to search with MCTSnets. arXiv preprint arXiv:1802.04697.
Gulcehre, C., Denil, M., Malinowski, M., Razavi, A., Pascanu, R., Hermann, K. M., Battaglia, P.,
Bapst, V., Raposo, D., Santoro, A., and de Freitas, N. (2018). Hyperbolic attention networks.
Hamaguchi, T., Oiwa, H., Shimbo, M., and Matsumoto, Y. (2017). Knowledge transfer for out-of-
knowledge-base entities: A graph neural network approach. In Proceedings of the International
Joint Conference on Artificial Intelligence (IJCAI).
Hamilton, W., Ying, Z., and Leskovec, J. (2017). Inductive representation learning on large graphs.
In Advances in Neural Information Processing Systems, pages 1025–1035.
Hamrick, J., Allen, K., Bapst, V., Zhu, T., McKee, K., Tenenbaum, J., and Battaglia, P. (2018).
Relational inductive bias for physical construction in humans and machines. In Proceedings of the
40th Annual Conference of the Cognitive Science Society.
Hamrick, J. B., Ballard, A. J., Pascanu, R., Vinyals, O., Heess, N., and Battaglia, P. W. (2017).
Metacontrol for adaptive imagination-based optimization. In Proceedings of the International
Conference on Learning Representations (ICLR).
Hartford, J., Graham, D. R., Leyton-Brown, K., and Ravanbakhsh, S. (2018). Deep models of
interactions across sets. arXiv preprint arXiv:1803.02879.
Hay, N., Stark, M., Schlegel, A., Wendelken, C., Park, D., Purdy, E., Silver, T., Phoenix, D. S.,
and George, D. (2018). Behavior is everything–towards representing concepts with sensorimotor
contingencies. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI).
Henaff, M., Bruna, J., and LeCun, Y. (2015). Deep convolutional networks on graph-structured
data. arXiv preprint arXiv:1506.05163.
Hinton, G. E. (1990). Mapping part-whole hierarchies into connectionist networks. Artificial
Intelligence, 46(1-2):47–75.
Hjort, N. L., Holmes, C., Müller, P., and Walker, S. G. (2010). Bayesian Nonparametrics. Cambridge
University Press.
Hoshen, Y. (2017). Vain: Attentional multi-agent predictive modeling. In Advances in Neural
Information Processing Systems, pages 2698–2708.
Hu, H., Gu, J., Zhang, Z., Dai, J., and Wei, Y. (2017). Relation networks for object detection.
29
Hudson, D. A. and Manning, C. D. (2018). Compositional attention networks for machine reasoning.
In Proceedings of the International Conference on Learning Representations (ICLR).
Humboldt, W. (1999/1836). On Language: On the diversity of human language construction and its
influence on the mental development of the human species. Cambridge University Press.
Hummel, J. E. and Holyoak, K. J. (2003). A symbolic-connectionist theory of relational inference

and generalization. Psychological Review, 110(2):220.
Ioffe, S. and Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing
internal covariate shift. In Proceedings of the 32nd International Conference on International
Conference on Machine Learning (ICML).
Johnson, D. D. (2017). Learning graphical state transitions. Proceedings of the International

Joulin, A. and Mikolov, T. (2015). Inferring algorithmic patterns with stack-augmented recurrent
nets. In Advances in Neural Information Processing Systems, pages 190–198.
Kansky, K., Silver, T., Mély, D. A., Eldawy, M., Lázaro-Gredilla, M., Lou, X., Dorfman, N., Sidor,
S., Phoenix, S., and George, D. (2017). Schema networks: Zero-shot transfer with a generative
causal model of intuitive physics. In Proceedings of the International Conference on Machine
Learning (ICML).
Kearnes, S., McCloskey, K., Berndl, M., Pande, V., and Riley, P. (2016). Molecular graph
convolutions: moving beyond fingerprints. Journal of computer-aided molecular design, 30(8):595–
608.
Kemp, C. and Tenenbaum, J. B. (2008). The discovery of structural form. Proceedings of the
National Academy of Sciences, 105(31):10687–10692.
Kipf, T., Fetaya, E., Wang, K.-C., Welling, M., and Zemel, R. (2018). Neural relational inference
for interacting systems. In Proceedings of the International Conference on Machine Learning
(ICML).
Kipf, T. N. and Welling, M. (2017). Semi-supervised classification with graph convolutional networks.
Koller, D. and Friedman, N. (2009). Probabilistic Graphical Models: Principles and Techniques.
MIT press.
Kondor, R., Son, H. T., Pan, H., Anderson, B., and Trivedi, S. (2018). Covariant compositional
networks for learning graphs. arXiv preprint arXiv:1801.02144.
Kondor, R. and Trivedi, S. (2018). On the generalization of equivariance and convolution in neural
networks to the action of compact groups. arXiv preprint arXiv:1802.03690.
Konidaris, G., Kaelbling, L. P., and Lozano-Perez, T. (2018). From skills to symbols: Learning
symbolic representations for abstract high-level planning. Journal of Artificial Intelligence
Research, 61:215–289.
Kool, W. and Welling, M. (2018). Attention solves your TSP. arXiv preprint arXiv:1803.08475.
30
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). ImageNet classification with deep convolu-
tional neural networks. In Advances in Neural Information Processing Systems, pages 1097–1105.
Kurach, K., Andrychowicz, M., and Sutskever, I. (2016). Neural random-access machines. In
Proceedings of the International Conference on Learning Representations (ICLR).
Lake, B. M. and Baroni, M. (2018). Still not systematic after all these years: On the compositional
skills of sequence-to-sequence recurrent networks. In Proceedings of the International Conference
on Machine Learning (ICML).
Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. (2015). Human-level concept learning
through probabilistic program induction. Science, 350(6266):1332–1338.
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., and Gershman, S. J. (2017). Building machines that
learn and think like people. Behavioral and Brain Sciences, 40.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521(7553):436.
LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., and Jackel,
L. D. (1989). Backpropagation applied to handwritten zip code recognition. Neural computation,
1(4):541–551.
Li, Y., Tarlow, D., Brockschmidt, M., and Zemel, R. (2016). Gated graph sequence neural networks.
Li, Y., Vinyals, O., Dyer, C., Pascanu, R., and Battaglia, P. (2018). Learning deep generative
models of graphs. In Workshops at the International Conference on Learning Representations
(ICLR).
Li, Y., Yu, R., Shahabi, C., and Liu, Y. (2017). Diffusion convolutional recurrent neural network:
Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926.
Lin, Z., Feng, M., Santos, C. N. d., Yu, M., Xiang, B., Zhou, B., and Bengio, Y. (2017). A structured
self-attentive sentence embedding. In Proceedings of the International Conference on Learning
Representations (ICLR).
Liu, H., Simonyan, K., Vinyals, O., Fernando, C., and Kavukcuoglu, K. (2018). Hierarchical
representations for efficient architecture search. In Proceedings of the International Conference on
Luong, M.-T., Pham, H., and Manning, C. D. (2015). Effective approaches to attention-based neural
machine translation. arXiv preprint arXiv:1508.04025.
Marcus, G. (2001). The algebraic mind.
Marcus, G. (2018a). Deep learning: A critical appraisal. arXiv preprint arXiv:1801.00631.
Marcus, G. (2018b). Innateness, alphazero, and artificial intelligence. arXiv preprint

arXiv:1801.05667.
McClelland, J. L. (1994). The interaction of nature and nurture in development: A parallel

distributed processing perspective. International perspectives on psychological science, 1:57–88.
31
McClelland, J. L. and Rumelhart, D. E. (1981). An interactive activation model of context effects
in letter perception: I. an account of basic findings. Psychological Review, 88(5):375.
Mikolov, T., Yih, W.-t., and Zweig, G. (2013). Linguistic regularities in continuous space word
representations. In Proceedings of the 2013 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies, pages 746–751.
Mitchell, T. M. (1980). The need for biases in learning generalizations. Department of Computer
Science, Laboratory for Computer Science Research, Rutgers Univ. New Jersey.
Mnih, V., Heess, N., Graves, A., et al. (2014). Recurrent models of visual attention. In Advances in
neural information processing systems, pages 2204–2212.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A.,
Riedmiller, M., Fidjeland, A. K., Ostrovski, G., et al. (2015). Human-level control through deep
reinforcement learning. Nature, 518(7540):529.
Monti, F., Boscaini, D., Masci, J., Rodola, E., Svoboda, J., and Bronstein, M. M. (2017). Geometric
deep learning on graphs and manifolds using mixture model cnns. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (CVPR).
Moravčı́k, M., Schmid, M., Burch, N., Lisỳ, V., Morrill, D., Bard, N., Davis, T., Waugh, K.,
Johanson, M., and Bowling, M. (2017). Deepstack: Expert-level artificial intelligence in heads-up
no-limit poker. Science, 356(6337):508–513.
Narayanan, A., Chandramohan, M., Chen, L., Liu, Y., and Saminathan, S. (2016). subgraph2vec:
Learning distributed representations of rooted sub-graphs from large graphs. In Workshops at the
20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.
Narayanan, A., Chandramohan, M., Venkatesan, R., Chen, L., Liu, Y., and Jaiswal, S. (2017).
graph2vec: Learning distributed representations of graphs. arXiv preprint arXiv:1707.05005.
Navon, D. (1977). Forest before trees: The precedence of global features in visual perception.
Cognitive Psychology, 9(3):353–383.
Niepert, M., Ahmed, M., and Kutzkov, K. (2016). Learning convolutional neural networks for
graphs. In Proceedings of the International Conference on Machine Learning (ICML), pages
2014–2023.
Nilsson, N. J. and Fikes, R. E. (1970). Strips: A new approach to the application of theorem proving
to problem solving. Technical report, SRI International, Menlo Park, CA Artificial Intelligence
Center.
Nowak, A., Villar, S., Bandeira, A. S., and Bruna, J. (2017). A note on learning algorithms for
quadratic assignment with graph neural networks. In Proceedings of the Principled Approaches to
Deep Learning Workshop (PADL) at the International Conference of Machine Learning (ICML).
Nowak, M. A. (2006). Five rules for the evolution of cooperation. science, 314(5805):1560–1563.
Ohtsuki, H., Hauert, C., Lieberman, E., and Nowak, M. A. (2006). A simple rule for the evolution
of cooperation on graphs and social networks. Nature, 441(7092):502.
32
Oñoro-Rubio, D., Niepert, M., Garcı́a-Durán, A., González-Sánchez, R., and López-Sastre,
R. J. (2017). Representation learning for visual-relational knowledge graphs. arXiv preprint
arXiv:1709.02314.
Parisotto, E., Mohamed, A.-r., Singh, R., Li, L., Zhou, D., and Kohli, P. (2017). Neuro-symbolic
program synthesis. In Proceedings of the International Conference on Learning Representations
(ICLR).
Pascanu, R., Li, Y., Vinyals, O., Heess, N., Buesing, L., Racanière, S., Reichert, D., Weber, T.,
Wierstra, D., and Battaglia, P. (2017). Learning model-based planning from scratch. arXiv
preprint arXiv:1707.06170.
Pearl, J. (1986). Fusion, propagation, and structuring in belief networks. Artificial intelligence,
29(3):241–288.
Pearl, J. (1988). Probabilistic reasoning in intelligent systems: Networks of plausible inference.
Morgan Kaufmann.
Pearl, J. (2009). Causality: Models, Reasoning and Inference. Cambridge University Press, New
York, NY, USA, 2nd edition.
Pearl, J. (2018). Theoretical impediments to machine learning with seven sparks from the causal
revolution. arXiv preprint arXiv:1801.04016.
Pennington, J., Socher, R., and Manning, C. (2014). Glove: Global vectors for word representation.
In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP),
pages 1532–1543.
Perozzi, B., Al-Rfou, R., and Skiena, S. (2014). Deepwalk: Online learning of social representations.
In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and
data mining, pages 701–710. ACM.
Pevnỳ, T. and Somol, P. (2017). Using neural network formalism to solve multiple-instance problems.
In International Symposium on Neural Networks, pages 135–142. Springer.
Pinker, S. and Prince, A. (1988). On language and connectionism: Analysis of a parallel distributed
processing model of language acquisition. Cognition, 28(1-2):73–193.
Plate, T. A. (1995). Holographic reduced representations. IEEE Transactions on Neural Networks,
6(3):623–641.
Plaut, D. C., McClelland, J. L., Seidenberg, M. S., and Patterson, K. (1996). Understanding normal
and impaired word reading: computational principles in quasi-regular domains. Psychological
Review, 103(1):56.
Pollack, J. B. (1990). Recursive distributed representations. Artificial Intelligence, 46(1-2):77–105.
Qi, C. R., Su, H., Mo, K., and Guibas, L. J. (2017). Pointnet: Deep learning on point sets for 3d
classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR).
Raposo, D., Santoro, A., Barrett, D., Pascanu, R., Lillicrap, T., and Battaglia, P. (2017). Discovering
objects and their relations from entangled scene representations. In Workshops at the International
33
Reed, S. and De Freitas, N. (2016). Neural programmer-interpreters. In Proceedings of the
International Conference on Learning Representations (ICLR).
Ritchie, D., Horsfall, P., and Goodman, N. D. (2016). Deep amortized inference for probabilistic
programs. arXiv preprint arXiv:1610.05735.
Rosenblatt, F. (1961). Principles of neurodynamics. perceptrons and the theory of brain mechanisms.
Technical report, Cornell Aeronautical Lab Inc., Buffalo, NY.
Rumelhart, D. E., McClelland, J. L., Group, P. R., et al. (1987). Parallel Distributed Processing,
volume 1. MIT Press.
Russell, S. J. and Norvig, P. (2009). Artificial Intelligence: A Modern Approach (3rd Edition).
Pearson.
Sabour, S., Frosst, N., and Hinton, G. E. (2017). Dynamic routing between capsules. In Advances
in Neural Information Processing Systems, pages 3859–3869.
Sanchez-Gonzalez, A., Heess, N., Springenberg, J. T., Merel, J., Riedmiller, M., Hadsell, R., and
Battaglia, P. (2018). Graph networks as learnable physics engines for inference and control. In
Proceedings of the 35th International Conference on Machine Learning (ICLR).
Santoro, A., Raposo, D., Barrett, D. G., Malinowski, M., Pascanu, R., Battaglia, P., and Lillicrap,
T. (2017). A simple neural network module for relational reasoning. In Advances in Neural
Information Processing Systems.
Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. (2009a). Computational
capabilities of graph neural networks. IEEE Transactions on Neural Networks, 20(1):81–102.
Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. (2009b). The graph
neural network model. IEEE Transactions on Neural Networks, 20(1):61–80.
Scarselli, F., Yong, S. L., Gori, M., Hagenbuchner, M., Tsoi, A. C., and Maggini, M. (2005). Graph
neural networks for ranking web pages. In Proceedings of the 2005 IEEE/WIC/ACM International
Conference on Web Intelligence, pages 666–672. IEEE.
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural Networks, 61:85–117.
Selsam, D., Lamm, M., Bunz, B., Liang, P., de Moura, L., and Dill, D. L. (2018). Learning a sat
solver from single-bit supervision. arXiv preprint arXiv:1802.03685.
Shalev-Shwartz, S., Shamir, O., and Shammah, S. (2017). Failures of gradient-based deep learning.
Shaw, P., Uszkoreit, J., and Vaswani, A. (2018). Self-attention with relative position representations.
In Proceedings of the 16th Annual Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies.
Shervashidze, N., Schweitzer, P., Leeuwen, E. J. v., Mehlhorn, K., and Borgwardt, K. M. (2011).
Weisfeiler-lehman graph kernels. Journal of Machine Learning Research, 12(Sep):2539–2561.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J.,
Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. (2016). Mastering the game of go with
deep neural networks and tree search. Nature, 529(7587):484–489.
34
Smolensky, P. (1990). Tensor product variable binding and the representation of symbolic structures
in connectionist systems. Artificial Intelligence, 46(1-2):159–216.
Socher, R., Huval, B., Manning, C. D., and Ng, A. Y. (2012). Semantic compositionality through
recursive matrix-vector spaces. In Proceedings of the Joint Conference on Empirical Methods in
Natural Language Processing (EMNLP) and Computational Natural Language Learning (CNLL),
pages 1201–1211. Association for Computational Linguistics.
Socher, R., Lin, C. C., Manning, C., and Ng, A. Y. (2011a). Parsing natural scenes and natural
language with recursive neural networks. In Proceedings of the 28th International Conference on
Machine Learning (ICML), pages 129–136.
Socher, R., Pennington, J., Huang, E. H., Ng, A. Y., and Manning, C. D. (2011b). Semi-supervised
recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference
on Empirical Methods in Natural Language Processing (EMNLP), pages 151–161. Association for
Computational Linguistics.
Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A., and Potts, C. (2013).
Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of
the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1631–1642.
Spelke, E. S., Breinlinger, K., Macomber, J., and Jacobson, K. (1992). Origins of knowledge.
Psychological review, 99(4):605.
Spelke, E. S. and Kinzler, K. D. (2007). Core knowledge. Developmental Science, 10(1):89–96.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. (2014). Dropout:
A simple way to prevent neural networks from overfitting. The Journal of Machine Learning
Research, 15(1):1929–1958.
Sukhbaatar, S., Fergus, R., et al. (2016). Learning multiagent communication with backpropagation.
Sukhbaatar, S., Weston, J., Fergus, R., et al. (2015). End-to-end memory networks. In Advances in
Neural Information Processing Systems, pages 2440–2448.
Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Sequence to sequence learning with neural networks.
Szegedy, C., Ioffe, S., Vanhoucke, V., and Alemi, A. A. (2017). Inception-v4, inception-resnet
and the impact of residual connections on learning. In Proceedings of the AAAI Conference on
Artificial Intelligence (AAAI), volume 4, page 12.
Tai, K. S., Socher, R., and Manning, C. D. (2015). Improved semantic representations from
tree-structured long short-term memory networks. In Proceedings of the Annual Meeting of the
Association for Computational Linguistics (ACL).
Tang, J., Qu, M., Wang, M., Zhang, M., Yan, J., and Mei, Q. (2015). Line: Large-scale information
network embedding. In Proceedings of the 24th International Conference on World Wide Web,
pages 1067–1077. International World Wide Web Conferences Steering Committee.
Tenenbaum, J. B., Griffiths, T. L., and Kemp, C. (2006). Theory-based bayesian models of inductive
learning and reasoning. Trends in Cognitive Sciences, 10(7):309–318.
35
Tenenbaum, J. B., Kemp, C., Griffiths, T. L., and Goodman, N. D. (2011). How to grow a mind:
Statistics, structure, and abstraction. Science, 331(6022):1279–1285.
Toyer, S., Trevizan, F., Thiebaux, S., and Xie, L. (2017). Action schema networks: Generalised
policies with deep learning. In Proceedings of the AAAI Conference on Artificial Intelligence
(AAAI).
Ullman, T. D., Spelke, E., Battaglia, P., and Tenenbaum, J. B. (2017). Mind games: Game engines
as an architecture for intuitive physics. Trends in Cognitive Sciences, 21(9):649–665.
van Steenkiste, S., Chang, M., Greff, K., and Schmidhuber, J. (2018). Relational neural expectation
maximization: Unsupervised discovery of objects and their interactions. Proceedings of the
International Conference on Learning Representations (ICLR).
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and
Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing
Systems.
Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., and Bengio, Y. (2018). Graph
attention networks. In Proceedings of the International Conference on Learning Representations
(ICLR).
Wang, J. X., Kurth-Nelson, Z., Kumaran, D., Tirumala, D., Soyer, H., Leibo, J. Z., Hassabis, D.,
and Botvinick, M. (2018a). Prefrontal cortex as a meta-reinforcement learning system. Nature
neuroscience, page 1.
Wang, J. X., Kurth-Nelson, Z., Tirumala, D., Soyer, H., Leibo, J. Z., Munos, R., Blundell, C.,
Kumaran, D., and Botvinick, M. (2016). Learning to reinforcement learn. arXiv preprint
arXiv:1611.05763.
Wang, T., Liao, R., Ba, J., and Fidler, S. (2018b). Nervenet: Learning structured policy with graph
neural networks. In Proceedings of the International Conference on Learning Representations
(ICLR).
Wang, X., Girshick, R., Gupta, A., and He, K. (2018c). Non-local neural networks. In Proceedings
of the Conference on Computer Vision and Pattern Recognition (CVPR).
Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M., and Solomon, J. M. (2018d). Dynamic
graph cnn for learning on point clouds. arXiv preprint arXiv:1801.07829.
Watters, N., Zoran, D., Weber, T., Battaglia, P., Pascanu, R., and Tacchetti, A. (2017). Visual
interaction networks: Learning a physics simulator from video. In Advances in Neural Information
Processing Systems, pages 4542–4550.
Wu, J., Lu, E., Kohli, P., Freeman, B., and Tenenbaum, J. (2017). Learning to see physics via
visual de-animation. In Advances in Neural Information Processing Systems, pages 152–163.
Yoon, K., Liao, R., Xiong, Y., Zhang, L., Fetaya, E., Urtasun, R., Zemel, R., and Pitkow, X.
(2018). Inference in probabilistic graphical models by graph neural networks. In Workshops at
the International Conference on Learning Representations (ICLR).
You, J., Ying, R., Ren, X., Hamilton, W. L., and Leskovec, J. (2018). GraphRNN: A deep generative
model for graphs. arXiv preprint arXiv:1802.08773.
36
Yuille, A. L. and Liu, C. (2018). Deep nets: What have they ever done for vision? arXiv preprint
arXiv:1805.04025.
Zaheer, M., Kottur, S., Ravanbakhsh, S., Poczos, B., Salakhutdinov, R. R., and Smola, A. J. (2017).
Deep sets. In Advances in Neural Information Processing Systems, pages 3394–3404.
Zambaldi, V., Raposo, D., Santoro, A., Bapst, V., Li, Y., Babuschkin, I., Tuyls, K., Reichert, D.,
Lillicrap, T., Lockhart, E., Shanahan, M., Langston, V., Pascanu, R., Botvinick, M., Vinyals, O.,
and Battaglia, P. (2018). Relational deep reinforcement learning. arXiv preprint arXiv.
Zhang, A., Lerer, A., Sukhbaatar, S., Fergus, R., and Szlam, A. (2018). Composable planning with
attributes. arXiv preprint arXiv:1803.00512.
Zügner, D., Akbarnejad, A., and Günnemann, S. (2018). Adversarial Attacks on Neural Networks
for Graph Data. arXiv preprint arXiv:1805.07984.
37
Appendix: Formulations of additional models
In this appendix we give more examples of how published networks can fit in the frame defined by
Equation 1.
Interaction networks
Interaction Networks (Battaglia et al., 2016; Watters et al., 2017) and the Neural Physics Engine
Chang et al. (2017) use a full GN but for the absence of the global to update the edge properties:
φe (ek , vrk , vsk , u) := f e (ek , vrk , vsk ) = NNe ([ek , vrk , vsk ])

X
ρe→v Ei0 := e0k

=
{k: rk =i}
That work also included an extension to the above formulation which output global, rather than
per-node, predictions:
φe (ek , vrk , vsk , u) := f e (ek , vrk , vsk ) = NNe ([ek , vrk , vsk ])

φu ē0 , v̄0 , u := f u v̄0 , u = NNu [v̄0 , u]

X
ρv→g V 0 := vi0

=
i
Non-pairwise interactions
Gated Graph Sequence Neural Networks (GGS-NN) (Li et al., 2016) use a slightly generalized
formulation where each edge has an attached type tk ∈ {1, .., T }, and the updates are:
φe ((ek , tk ) , vrk , vsk , u) := f e (ek , vsk ) = NNe,tk (vsk )

φv ē0i , vi , u := f v ē0i , vi = NNv [ē0i , vi ]

X
ρe→v Ei0 := e0k

=
{k: rk =i}
These updates are applied recurrently (the NNv is a GRU (Cho et al., 2014)), followed by a global
decoder which computes a weighted sum of embedded final node states. Here each NNe,tk is a neural
network with specific parameters.
CommNet (Sukhbaatar et al., 2016) (in the slightly more general form described by (Hoshen,
2017)) uses:
φe (ek , vrk , vsk , u) := f e (vsk ) = NNe (vsk )

φv ē0i , vi , u := f v ē0i , vi = NNv [ē0i , NNv0 (vi )]

1 X
ρe→v Ei0 := e0k

= 0
|Ei |
{k: rk =i}
38
Attention-based approaches
The various attention-based approaches use a φe which is factored into a scalar pairwise-interaction
function which returns the unnormalized attention term, denoted αe (vrk , vsk ) = a0k , and a vector-
valued non-pairwise term, denoted β e (vsk ) = b0k ,
φe (ek , vrk , vsk , u) := f e (vrk , vsk ) = (αe (vrk , vsk ) , β e (vsk )) = (a0k , b0k ) = e0k
The single-headed self-attention (SA) in the Transformer architecture (Vaswani et al., 2017),
implements the non-local formulation as:
αe (vrk , vsk ) = exp (NNαquery (vrk )| · NNαkey (vsk ))

β e (vsk ) = NNβ (vsk )
v
ē0i , vi , u := f v ē0i = NNv ē0i

φ
where NNαquery , NNαkey , and NNβ are again neural network functions with different parameters and
possibly different architectures. They also use a multi-headed version which computes Nh parallel
ē0h
i using different NNαh
query , NN key , NNβ , where h indexes the different parameters. These are
αh h
v
passed to f and concatenated:

0Nh
f v {ē0h 01
i }h=1...N h = NNv [ēi , . . . , ēi ]
Vertex Attention Interaction Networks (Hoshen, 2017) are very similar to single-headed SA,
but use Euclidean distance for the attentional similarity metric, with shared parameters across the
attention inputs’ embeddings, and also use the input node feature in the node update function,
αe (vrk , vsk ) = exp −kNNα (vrk ) − NNα (vsk ) k2

β e (vsk ) = NNβ (vsk )

v
ē0i , vi , u := f v ē0i = NNv [ē0i , vi ]

φ
Graph Attention Networks (Veličković et al., 2018) are also similar to multi-headed SA, but use
a neural network as the attentional similarity metric, with shared parameters across the attention
inputs’ embeddings:
αe (vrk , vsk ) = exp (NNα0 ([NNα (vrk ) , NNα (vsk )))

e
β (vsk ) = NNβ (vsk )

0Nh
φv ē0i , vi , u := f v {ē0h 01

i }h=1...Nh = NNv [ēi , . . . , ēi ]
Stretching beyond the specific non-local formulation, Shaw et al. (2018) extended multi-headed
SA with relative position encodings. “Relative” refers to an encoding of the spatial distance between
nodes in a sequence or other signal in a metric space. This can be expressed in GN language as an
edge attribute ek , and replacing the β e (vsk ) from multi-headed SA above with:
β e (ek , vsk ) = NNe (vsk ) + ek
39
Belief Propagation embeddings
Finally, we briefly summarize how the general “structure2vec” algorithm of Dai et al. (2016) can fit
into our framework. In order to do so, we need to slightly modify our main Equation 1, i.e.:
X
¯k = ρ {el }sl =rk := el
rl 6=sk rl =sk
sl 6=rk
e0k = φe (¯
k ) := f (¯
k ) = NN(¯
k )
X
ē0i = ρ {e0k }rk =i

:= ek
{k: rk =i}
vi0 v
ē0i := f (ē0i ) = NN(ē0i )

=φ
Edges’ features now takes the meaning of “message” between their receiver and sender; note that
there is only one set of parameters to learn for both the edges and nodes updates.
40

Relational Inductive Biases, Deep Learning, and Graph Networks

Uploaded by

Copyright:

Available Formats

You might also like

Relational Inductive Biases, Deep Learning, and Graph Networks

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Relational Inductive Biases, Deep Learning, and Graph Networks

Uploaded by

Copyright:

Available Formats

Relational inductive biases, deep learning, and graph networks

Peter W. Battaglia1∗, Jessica B. Hamrick1 , Victor Bapst1 ,

Victoria Langston1 , Chris Dyer1 , Nicolas Heess1 ,

1 DeepMind; 2 Google Brain; 3 MIT; 4 University of Edinburgh

As an illustrative example of relational reasoning in machine learning, graphical models (Pearl,

2 Relational inductive biases

2.1 Relational inductive biases in standard deep learning building blocks

2.1.2 Convolutional layers

2.1.3 Recurrent layers

2.2 Computations over sets and graphs

(c) n-body System (d) Rigid Body System

3.2 Graph network (GN) block

Directed : one-way edges, from a “sender” node to a “receiver” node.

3.2.1 Definition of “graph”

3.2.2 Internal structure of a GN block

A GN block contains three “update” functions, φ, and three “aggregation” functions, ρ,

e0k = φe (ek , vrk , vsk , u) ē0i = ρe→v Ei0

vi0 = φv ē0i , vi , u ē0 = ρe→u E 0

3.2.3 Computational steps within a GN block

(a) Edge update (b) Node update (c) Global update

3.2.4 Relational inductive biases in graph networks

4.1 Flexible representations

4.1.2 Graph structure

4.2 Configurable within-block structure

φu ē0 , v̄0 , u := f u ē0 , v̄0 , u = NNu [ē0 , v̄0 , u]

Edge block Node block Global block

(a) Full GN block (b) Independent recurrent block

(c) Message-passing neural network (d) Non- oca neura network

(e) Re at on network (f) Deep set

4.2.1 Message-passing neural network (MPNN)

4.2.2 Non-local neural networks (NLNN)

◦ their f plays the role of the above α,

4.2.3 Other graph network variants

φe (ek , vrk , vsk , u) := f e (vrk , vsk ) = NNe ([vrk , vsk ])

φv (ēi , vi , u) := f v (vi , u) = NNv ([vi , u])

GNcore Gthid1 GNcore Gthid

(a) Composition of GN blocks (b) Encode-process-decode (c) Recurrent GN arch tecture

4.3 Composable multi-block architectures

4.4 Implementing graph networks in code

Shortest path demo: tinyurl.com/gn-shortest-path-demo

Sort demo: tinyurl.com/gn-sort-demo

Physics demo: tinyurl.com/gn-physics-demo

5.1 Combinatorial generalization in graph networks

5.2 Limitations of graph networks

5.3 Open questions

5.4 Integrative approaches for learning and structure

Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89(4):369.

Botvinick, M. M. (2008). Hierarchical models of behavior and prefrontal function. Trends in

Chomsky, N. (1957). Syntactic Structures. Mouton & Co.

Chomsky, N. (1965). Aspects of the Theory of Syntax. MIT Press.

Cohen, T. and Welling, M. (2016). Group equivariant convolutional networks. In International

Craik, K. J. W. (1943). The Nature of Explanation. Cambridge University Press.

Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2):179–211.

Elman, J. L. (1991). Distributed representations, simple recurrent networks, and grammatical

Fodor, J. A. (1975). The Language of Thought. Harvard University Press.

Fodor, J. A. and Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical

Hummel, J. E. and Holyoak, K. J. (2003). A symbolic-connectionist theory of relational inference

Johnson, D. D. (2017). Learning graphical state transitions. Proceedings of the International