Week 7

You might also like

Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 6

Week 6

Organic Minds –
Brains reasoning and rationality – decision theory, Bayesian reasoning, heuristics
Describe human behaviour at different level of analysis – neural, behavioral/mechanistic and
computational levels
Optimality of human decision making

Visual Guide to Bayesian Thinking


- Tom is shy example (math PhD and business school students)
- Robbers and honest probability example
- Bob is jealous and complains of Alice example
- Bayesian thinking: remember priors; imagine theory’s wrong. Would it world look
different; update incrementally
- Meditation is fake v. real

Daniel Wolpert: The real reason for brains


Why do we and animals have brains? – Produce adaptable and complex movements
Communication all through movements
Example: Sea Squirts implant on rock – digest brain and nervous system after locking in
cause not needed
Chess example – all possible moves and algorithms – optimal solution
Pour water robot
Cup stacker example – command to give muscle movement
Random noise – corrupts signal – sensory feedback and movement command along with
ambiguous feedback of task

Bayesian decision theory - 2 sources of information


1. Data e.g. sensory input
2. Prior knowledge e.g. memory
3. Leads to Beliefs
Mathematical computation provided
Tennis example – vision and auditory information + variability and other information like
prior information about tennis
Tickling example and escalation of force and Tit for tat example
Justin Gardner Video on Perception
Forward direction hierarchical sensory processing – sensory transduction like eyes and other
stimulations; thalamic processing; primary sensory cortex; categorical representations
- Natural image categorization example in machine learning as example
Backward direction Bayesian inference – sensory responses are inherently ambiguous and
noisy – what causes that sensory action // priors can disambiguate to infer properties of world
(prior, posteriors best guess and sensory evidence likelihood)
Example: In the 50's and 60's, David Hubel and Torsten Wiesel measured from visual neurons
in the cat brain – paralyzed cat
Electrical signals on axon – dendrites – action firing
LGN (Lateral Geniculate Nucleus) – receptive neuron field – antagonism process
Both on and off-Centre types
Orientation selective simple receptive field
Fusiform gyrus - face area categorization
Spector and Parvizi – electrical stimulation of fusiform face area
Canonical circuit – image categorization – deep convolution neural networks

Bayesian inference
Example of illusion – A and B box – luminance – automatic computation you are doing when
seeing it
Bumps sticking inside or outside ambiguity - example – gets data from where sun is located
3d shapes like foot in sand
Hermann Von Helmholtz – unconscious inference theory – use prior information and
combine with sensory evidence in mathematically optimal way to get a posterior distribution
that tell you what the best guess is
Example: tennis ball landing

Decision Theory video by RA Briggs


Decision – act with best outcome – decisions based out of uncertainty conditions – 2
questions asked: how do reasoning agents make decisions (descriptive); and how should
reasoning agents make these decisions (normative) = answer to this is expected utility
Expected utility = acts (things you can choose to do, in your control, and evaluate outcomes);
outcomes (results of acts – good or bad in themselves); and events (happening entirely
outside your control which combine with acts to produce outcomes) // states (events so
specific that together with any act you choose, they fully determine an outcome)
Example: making omelet 6 rotten case
Probabilities and Utilities
1. Events have probabilities. > Probability is open to interpretation (classical,
frequentist, subjectivist). Many decision theorists lean toward subjectivist
interpretations.
2. ▪ Outcomes have utilities. > (Real) numbers that measure how good an outcome is. >
Unit : utils > The higher the number, the better the outcome. > How informative are
these numbers? • The ordering matters. • The distances between numbers matter. The
absolute magnitudes don't matter.
Arguments that people either do Maximize expected utility or should maximize it based on
how interpret it
Why Do (Should) People Maximize Expected Utility? One type of argument : it's obvious. ▪
Another type of argument : it leaves you better off in the long run. ▪ A third type of
argument : representation. 1. If you are rational, your preferences will satisfy certain axioms.
2. If your preferences satisfy the axioms in premise 1, then you can will behave as if you are
maximizing expected utility. I 3. If you behave as if you are maximizing expected utility, then
you are maximizing expected utility. ▪ Representation theorems are mathematical arguments
that establish premise 2.
John Von Neurmann and Oskar Morgenstern axioms
Are People Bayesian reasoners video by Michael Frank
Uncertainty everywhere – future, weather, stock market, cycling, infants
Bayes is optimal way of reasoning – Bayesian reasoning about probability is normative –
gives best answers (IF prior data and hypothesis match what we care about)
Dr. Pangloss from Voltaire’s Candide example – have real problem and real solution in mind
Example 1: Prior neglect experiment by Daniel Kahneman and Amos Tversky (1972) – 2 cab
companies
Implicit Bayesian (faster, less conscious awareness, not always verbal) v. Explicit Bayesian
(slower, conscious, verbal)
Example 2: Bayesian perception - ■ Bayesian inference can be used to reason about uncertain
perceptual input Surprisingly, the human visual system often appears to make inferences in
this way Populations of neurons can compute in ways that sometimes look like combining
prior and likelihood
Example 3: Bayesian generalization of vegetables – Subordinate matches, basic level
matches, superordinate matches
Other examples: Goodman and Frank (Rational speech model); Gweon (sampling);
Tenenbaum
Example: Detective investigating a crime scene has some prior beliefs based on their
experience as a detective and the information gathered from previous cases similar to Visual
Guide to Bayesian Thinking’s video with Tom’s example. These prior beliefs could include
general patterns of criminal behavior, common motives, and potential suspects based on
their past criminal history. This is similar to - Hermann Von Helmholtz – unconscious
inference theory – use prior information and combine with sensory evidence in
mathematically optimal way to get a posterior distribution that tell you what the best guess is

new evidence at the crime scene, such as fingerprints, DNA samples, and witness
statements. constitutes the likelihood or current evidence in the Bayesian framework.
Each piece of evidence contributes to updating your prior beliefs. Daniel Wolpert’s
The real reason for brains video provides for Bayesian decision theory - 2 sources of
information 1. Data e.g. sensory input 2. Prior knowledge e.g. memory 3. Leads to
Beliefs Mathematical computation provided

using Bayesian thinking as Justin Gardner Video on Perception - Backward direction


Bayesian inference - analyze the new evidence in light of prior beliefs; combine prior
beliefs with the likelihood of the new evidence to compute the posterior probability
of different hypotheses, such as who the suspect might be or what happened during
the crime. gather more evidence and information, continuously update your beliefs.
the new evidence contradicts initial assumptions, might revise your prior beliefs
accordingly.

Example: Availability Heuristic


Behavior: I am trying to decide which mode of transportation to take for a trip. I start
thinking about various options and their potential risks, like driving a car, taking a train, or
flying in an airplane. Are People Bayesian reasoners video by Michael Frank shows that there
is Uncertainty everywhere. While considering the risks, I recall recent news stories about
airplane crashes, which are vivid and emotionally impactful.

Heuristic Bias/Principle: The availability heuristic is a cognitive bias where people tend to
judge the likelihood of an event based on how easily they can bring examples of such events
to mind. In this case, my judgment of the risk associated with flying is influenced by the ease
with which I can recall recent airplane crash news.
Outcome: Due to the availability heuristic bias, I might end up perceiving flying as a riskier
mode of transportation than it statistically is. Consequently, I might decide to avoid flying
and opt for a less safe mode of transportation, even though statistically, flying is one of the
safest ways to travel.

This example illustrates how the availability heuristic can lead to biases in decision-making.
Instead of objectively evaluating the actual risks associated with different modes of
transportation using Bayesian reasoning, my judgment is swayed by the ease of recall of
emotionally charged events. While Bayes is optimal way of reasoning – Bayesian reasoning
about probability is normative – gives best answers (IF prior data and hypothesis match what
we care about). In reality, being aware of the statistical probabilities and using Bayesian
reasoning to weigh the likelihood of accidents would be a more rational approach to making a
decision.

Example: A detective investigating a crime scene (in a web series) approaches the task with
valuable prior beliefs based on years of experience and insights gained from previous cases,
much like the approach described in the Visual Guide to Bayesian Thinking's video featuring
Tom's example. These prior beliefs encompass general patterns of criminal behavior,
common motives, and potential suspects based on historical criminal data. This aligns with
Hermann Von Helmholtz's unconscious inference theory, which involves utilizing prior
information and integrating it with sensory evidence in a mathematically optimal manner to
form a posterior distribution, leading to the best possible inference about the new evidence
found at the crime scene. The new evidence includes fingerprints, DNA samples, and witness
statements, constituting the likelihood or current evidence in the Bayesian framework.

Drawing inspiration from Daniel Wolpert's video on the real reason for brains, which is
rooted in Bayesian decision theory, there are two essential sources of information: sensory
input (data) and memory (prior knowledge). These sources converge to shape beliefs through
mathematical computations. In a similar vein, Justin Gardner's Perception video highlights
the use of backward-direction Bayesian inference. It involves analyzing new evidence in light
of prior beliefs and combining both to compute the posterior probability of different
hypotheses, such as the suspect's identity or the sequence of events during the crime.
As the investigation progresses, the detective continuously gathers more evidence and
information, updating their beliefs accordingly. If the new evidence contradicts initial
assumptions, the detective remains open to revising their prior beliefs to refine their
understanding of the case.

You might also like