CS781:

A Quick Primer on
Abstract Interpretation for
Neural Networks

Supratik Chakraborty
IIT Bombay
 Notion of State in Neural
Network
x3 x6 x9
x12
x15 x17
x1

x4 x10

x7 x13 x16
x18
x2 x5 x11
x8 x14

State: (x1, x2, ... x18) in R18

State Change in Feed-Forward
Neural Network
x3 x6 x9

x1

x4

x7

x2 x5
x8

(x’1, x’2, ... x’i-1, x’i) = fi(x1, x2, ... xi-1), for i in {3, ..., 18 }
 State Change in Feed-Forward
NN as a sequence of instrns
x3
x1

x2

(x’1, x’2, x’3) = f3 (x1, x2);

(x’’1, x’’2, x’’3, x’’4) = f4 (x’1, x’2 , x’3);
…

NN computation: a sequence of state transitions

caused by seq of instructions
 Proving Property of a FF NN
x3
x1 x17
Pre- Post-
condition condition on
on (x1, x2) (x17, x18)
x18
x2

{Pre-condition on (x1, x2)}

(x’1, x’2, x’3) = f 3 (x1, x2);

(x’’1, x’’2, x’’3, x’’4)
= f4 (x’1, x’2 , x’3);
…
{Post-condition on (x17, x18) }
NN Computation as a State Transition
System

{Pre-condition on (x1, x2)}

(x’1, x’2, x’3) = f3 (x1, x2);

(x’’1, x’’2, x’’3, x’’4)
= f4 (x’1, x’2 , x’3);
…
{Post-condition on (x17, x18) }
Dealing with State Space Size

Infinite state space

Difficult to represent using state transition diagram

Can we still do some reasoning?
Concrete states

Solution: Use of abstraction

Naive view
●
Bunch sets of states together “intelligently”
●
Don't talk of individual states, talk of a representation of a set
of states
●
Transitions between state set representations

Granularity of reasoning shifted Abstract states

Extremely powerful general technique
●
Allows reasoning about large/infinite state spaces
A Generic View of Abstraction
Set of concrete states Set of abstract states
Abstraction (a)

Concretization (g)


Every subset of concrete states mapped to
unique abstract state

Desirable to capture containment relations

Transitions between state sets (abstract states)
 The Game Plan

C A
O a B
N S
C T
R R
E A
T C
E T Yes,
Pre-condition: g Proof
S S
T T
NN computation A A
as a T T
sequence of
state transitions
E E
S S No,
Counterexample
Post-condition:
Abstract analysis engine
 The Game Plan

C A
O a B
N S
C T
R R
E A
T C
E T Yes,
g Proof
How do we S choose the right
S abstraction?
Is there a method
T beyond domain
T expertise?
Can we learnA from errors in A
abstraction to build
T (refined) abstractions?
better T
E E
CanSrefinement be automated?
S No,
Counterexample

Abstract analysis engine

 The Game Plan

C A
Abstract state spaces
O can be
a infinite.
B
What can we do toN make abstractS
C
analysis practical? T
R R
Finite ascending
E chains A
what beyond?
T C
E T Yes,
g Proof
S S
T T
A A
T T
E E
S S No,
Counterexample

Abstract analysis engine

Desirable Properties of Abstraction
Set of concrete states Set of abstract states
Abstraction (a)

Concretization (g)


Suppose : subsets of concrete states
●
Any behaviour starting from can also happen starting from

●
If , we want this monotonicity in
behaviour in abstr state space too
●
Need ordering of abstract states, similar in spirit to
Structure of Concrete State Space


Set of concrete states: S

Concrete lattice C =

Powerset of S

Bottom
element
Partial order

Least upper bound Top element

Greatest lower bound

Structure of Abstract State Space


Abstract lattice A=

Abstraction function

●
Monotone: for all
●


Concretization function

●
Monotone: for all 2
●