Cognitive science#
Some notes on a computational perspective on cognitive science.
About this chapter
The chapter starts from the inverse problem of cognitive science: given the world (say, pixels), how does a mind form representations of it? Nativists build representations in, empiricists learn them, and constructivists treat children as scientists who revise theories. Probabilistic causal models make the constructivist view precise with Bayes’ rule, while the deep learning sections take the empiricist side: little built-in structure and lots of data.
Later sections cover language development (Skinner against Chomsky), theory of mind, and the classic AI view of representation from CS 188 and Russell & Norvig: physical symbol systems, heuristic search, logic, semantic networks and expert systems. The chapter ends with notes on Gödel, Escher, Bach and on cognitive biases from LessWrong.
Chapter outline: 8 sections
- Nativism, empiricismThree answers to how representations form.
- Probabilistic causal modelsBayesian models of cognition ask which priors would make people's inferences rational.
- Deep rl
- ObjectsNotes on visual concepts: adapting across datasets, explaining vision-and-language models, and learning categories.
- Language developmentThe central debate.
- Theory of mindTheory of mind is understanding other agents' mental states and how they shape actions.
- RepresentationsThese notes follow CS 188 and Russell & Norvig on how classic AI represented knowledge: as symbols manipulated by search (Newell and Simon's physical symbol system hypothesis), as logic, as semantic networks and frames, or statistically.
- Lesswrong1 interactiveA cognitive bias is a systematic error in judgment, like the bias of an estimator.
Nativism, empiricism#
Three answers to how representations form. Nativism builds them in, empiricism learns them from experience, and constructivism (Piaget, and later the theory theory) has children revise theories as scientists do. The contemporary versions follow, with their main proponents.
cogsci “inverse problem” - give world, how do we form representations
e.g. given pixels, how do we understand world?
historical cognitive development: how do we form representations?
nativism - plato - representation w/out learning
empiricism - aristotle - learning w/out representation
constructivism - jean piaget - cognitive development
types
assimilation - start with schema, which tells you how to act
accomodation - adjust schema based on what you see
equilibration - everything set
periods
sensorimotor
preoperational
concrete operational
formal operational
0 - 18 months
18 months - 5 years
5 years - adolescence
adolescence and beyond
problems
dev. is domain-specific and variable
when measured better, children show earlier competence
doesn’t specify learning methods
contemporary theories
nativism - core knowledge or modules
plato, descartes
chomsky - language acuquisition device
spelke, tenenbaum - core knowledge of domains
constraint nativism vs. starting-state nativism
empiricism - connectionism, dynamic systems, associationism
aristotle, david hume
behaviorism - led to rl
connectionists - mclelland, karmiloff-smith
dynamic systems - thelen and smith
emergentist appproach - complex behavior is emergent from simple neural properties
john locke
deep learning
constructivism - “the theory theory” (children are scientists), probabilistic models
carey, wellman, gelman, gopnik
structural features: abstract, coherent, causal, hierarchical
theories change in response to evidence
changes may take place at multiple levels
playing
perhaps bayesian learning
information-processing - memory changes over time etc.
siegler
socio-cultural influences - children learn from society / culture
vyogotsky
unanswered questions
search problem: how do children search through all possible hypotheses? sampling?
conceptual change problem: how is radical change in representations possible
counterfactual - relating to or expressing what has not happened or is not the case
e.g. If kangaroos had no tails, they would topple over
Probabilistic causal models#
Bayesian models of cognition ask which priors would make people’s inferences rational. They describe what is being computed, Marr’s computational level, rather than the mechanism.
abstract representations that provide computational accounts of causal inference
at marr’s highest level: computation
how much do we need to build in?
dna can build in things that are hard to learn
start with nothing built in, ex. deep learning (connectionism)
start with best possible learning algorithm and ask what you need to build in (bayesian modeling)
bayes rule: \(\overbrace{p(\theta \vert x)}^{\text{posterior}} = \frac{\overbrace{p(x\vert \theta)}^{\text{likelihood}} \overbrace{p(\theta)}^{\text{prior}}}{p(x)}\) where x is data and \(\theta\) are hypotheses
ask what priors explain people’s inferences
humans make very causal priors - restricts hypothesis space of possible bayesian networks
hierarchical model - get prior from something (e.g. know all bags contain same color)
what develops over time?
bayesian doesn’t really tell us this - just has probabilities evolve over time
real life we come up with new hypotheses
what representations do we use?
Deep rl#
DL basics#
Deep learning takes the empiricist side, and these notes ask how much inductive bias a network needs. The bullets argue that less built-in structure plus more data has worked better, with exceptions such as CNNs and LSTMs.
what does an NN do?
universal function approximator
feature learning view - learn features that separate classes linearly
bhand…saxe, ng 2010 compared statistics of primary cortical receptive fields and receptive field plasticity
vision + audio + somatosensory - statistics seem to line up with neuroscience
a single algorithm?
ex. seeing with your tongue
ex. blind people echolocating
ex. ferret optic nerve rewired to auditory system and they learn to still see
lots of inductive bias? or lack of bias?
gabor filters come from deep learning, ica, sparse coding, k-means - come from data, not necessarily algorithm
maybe this is only true for gabor filters though
bigger and simpler with residuals seems to work best
sequence tasks - lstm started being replaced with “attention is all you need” (attention is basically a dot product)
differentiable neural computer didn’t seem to work all that well - tried to build in too much
some inductive bias has worked: cnn, lstm
structure
structure is very dependent on optimization
depth seems to help
compositionality
increase in representational power?
information theoretic arguments
low-data regimes: few-shot recognition doesn’t work (see Lake, Salakhutidinov, Tenenbaum 2013)
one type of meta-learning for few-shot learning
split data set into tiny train / test blocks and learn to do few-shot learning on these
got really good at this (especially for omniglot)
deep learning: less structure + more data = better results
alternative view - not necessarily dl
very unconstrained function class + right learning algorithm + right supervision = better results
optimization methods master (is sgd “conveniently” Bayesian?)
ex. everything that works works because it’s Bayesian blog post
structure of natural dat amatters
source of supervision really matters
maybe function class doesn’t really matter
Supervision#
The source of supervision matters as much as the model. The notes compare unsupervised learning of \(p_\theta(x)\) (deep belief nets, autoencoders, VAEs, GANs), supervised and self-supervised learning, and reinforcement learning, where the only signal is a reward that cannot be differentiated.
unsupervised learning: learn \(p_\theta (x)\)
overview
pull out features/representation and use them for other tasks
recognize novel / unexpected events
hallucinate
can use all available data
often has principled, prob. interpretation
hard to use effectively
deep belief networks (~2006) - built on restricted Boltzmann machine (both a NN and a graphical model)
RBM = markov field with binary variables and connections between but not within layers
originally trained layerwise
properties
each neuron is (binary) random variable, so we can sample
hard to train
unsupervised gd (2012-2014) - sparse/denoising/bottleneck autoencoder
ex. if you have linear encoder/decoder with bottleneck, then you get pca
ex. denoising - blur image and ask it to learn to clean it up
learns something about the manifold that represents a class / the correct data
ex. variational autoencders (Kingma & Welling 2013)
network encoder x->z & decoder z->x
force z to be spherical Gaussian (ex. add noise to make it spherical Gaussian (with learned variance) - KL divergence regularizer between output and spherical Gaussian)
then, we can sample z from spherical Gaussian and decoder will give us nice looking x
ex. GANs (Goodfewllow 2014)
“implicit” density model
supervised learning
representations are surprisingly good
labels provide good semantic knowledge
simple backprop training
semi-supervised = self-supervised learning
ex. given patches, tell where they are placed relative to each other (context prediction doersch et al. 2015)
prediction supervised learning (time series)
ex. predict video frames from past video frames
reakes some artistry to figure out objective
-
non-differentiable reward function - doesn’t access to output supervision
supervised learning is a subset of this
3 types (often we do a combination of these together)
direct policy search
prediction of value functions
prediction of future states
can collect data by just exploring world
rl needs more generalization
want more exploration algorithms
train in more settings
ex. approximate bayesian experiment design
Objects#
Notes on visual concepts: adapting across datasets, explaining vision-and-language models, and learning categories. A recurring point is that the gains come from data. With enough data nearest neighbors works well, and human visual memory is large enough that the brain might work similarly.
deep deep cnns - not just a chain, combine different layers and maybe represent semantic concepts
iterative deep aggregation - generalizes skip connections - new layers for the skips
adaptive learning - too much dataset bias - imagenet things only work on imagenet
need to know about discrepancy between distributions
GAN - learn adversarial that decides whether a point comes from one domain or another
goal: want both domains to be separable with domains for only one
explainable vision and language agents
image captioning doesn’t mean it actually understands scene
now, people focused on visual question answering
maybe even provide interpretable explanation
preprogram 5 types of modules (ex. find, relate, count, …)
concept vs. category learning
except for NN, need negative examples
learning nouns
strong sampling - pick key examples instead of lots of random examples
bayesian learning can figure out what classes to generalize to
need to assume some ontology
it’s all about the data
face detection - made things really work
learning spectrum: extrapolation (low samples) -> interpolation (high samples - where we are)
extrapolation = linear reg. -> interpolation = nearest neighbor / neural network
nearest neighbors started to work better with more data
explainability doesn’t work if system is fundamentally complicated
natural world has too many examples for interpolation
brain doing nearest neighbors?
capacity of visual long term memory - standing 1973: 10k images, 83% recognition
clap if you see a repetition - you can really remember a lot
AB testing at end - check if you seaw things or not
novel objects: 92%, different examples of same thing: 88%, different states of same thing: 87%
we can’t do this with just textured images though
big boosts come from data, not necessarily the algorithm
word2vec works about same as nearest neighbors embedding
top-down vs bottom-up concept learning
problems with top-down / using labels
can’t ask computer about semantic stuff - it only sees pixels
cool ex. can see a chair by just looking at silhouette of person sitting in it
humans don’t categorize things in binary ways
solns: hierarchy, levels of categories…
used to have to do this (ex. books in a library), but computers don’t have to do this (ex. amazon)
still problematic
intransitivity - car seat is chair, chair is furniture, …
multiple category membership
“ontologies are overrated” blog post
this helps with interpretability, but might be worse overal
start from image (bottom) - use association not categorization
make graph - edges are associations
task: predict what’s behind some blurred out box in an image
Language development#
The central debate. Skinner’s behaviorism says children learn language through reinforcement. Chomsky answers that children say things they have never heard, so they must bring innate knowledge of grammar.
lots of children have language difficulties
language might be unique to humans
language is somewhat instinctual
learned via association (John Locke) or by reading people’s intentions (St. Augustine)
behaviorist approach (Skinner) - conditioning, children learn because of adult reinforcement
ex. bird turning to get reward
“meaning” - an association between stimulus and response
progressilve taught more complex utterances
noam chomsky’s critique of skinner review (1959)
parents correct truth of utterance, not its grammar
stimulus doesn’t completely determine response
children say things they’ve never heard (e.g. mommy eated the apple)
grammar isn’t a chain of associated words
chomsky: children have innate knowledge
introduces “cognitive revolution” - explain language you need
mental representations
rules that operate on those representations
figure out what is common between all languages
ex. verb agrees with subject
mostly part of speaker’s unconscious knowledge of language
focus on ideal speaker’s competence, not their performance
ex. question formation
poverty of the stimulus - children get this info which doesn’t seem to be in the stimulus
chomsky theory of universal grammar
fixed invariant structural principles because human brain is wired to understand them
nativists - children guided by universal grammar
constructivists - innate general learning mechanisms operate on experience
children use babbling
children use pre-linguistic communication for sharing attention
some children have crib speech - talking to themselves, seems systematic
children have U-curves - use ate, then eated, then back to ate
Theory of mind#
Theory of mind is understanding other agents’ mental states and how they shape actions. It is acquired quickly but in a long series of steps, on a similar trajectory worldwide, and it is impaired in autism.
Theory of mind = The ability to understand and take into account another individual’s mental state
theory of mind - human understanding of agents’ mental states and how they shape actions
understand intentional actions - like dots, understand evil, purpose
we can track this month by month
Theory of mind
is rapidly acquired in the normal case
is acquired in an extended series of developmental accomplishments
encompasses several basic insights that are acquired world-wide on a roughly similar trajectory (but not timetable), (4) requires considerable learning and development based on an infant set of specialized abilities to attend to and represent persons, (5) is severely impaired in autism, (6) is severely delayed in deaf children of hearing parents, and (6) results from but also contributes to specialized neural substrates associated with reasoning about agency, experience, and mind
Preschool Theory of Mind, Part 1: Universal Belief-Desire Understanding
Representations#
These notes follow CS 188 and Russell & Norvig on how classic AI represented knowledge: as symbols manipulated by search (Newell and Simon’s physical symbol system hypothesis), as logic, as semantic networks and frames, or statistically.
based on berkeley’s cs 188 course and russell & norvig’s “artificial intelligence” (3rd edition)
Intro#
AI - field of study which studies the goal of creating intelligence
intelligent agent - system that perceives its environment and takes actions that maximize its chances of success
expert task examples - medical diagnosis, equipment repair, computer configuration, financial planning
formal systems - use axioms and formal logic
ontologies - structuring knowledge in graph form
statistical methods
turing test - is human mind deterministic { turing1950computing }
chinese room argument - rebuts turing test { cite searle1980minds }
china brain - what if different people hit buttons to fire individual neurons
Polanyi’s paradox: “we can know more than we can tell”
Symbol search#
Newell and Simon’s view: intelligence is the manipulation of symbols, and a symbol system solves problems by heuristic search, generating and modifying symbol structures until one is a solution.
computer science - empirical inquiry
Symbols and physical symbol systems#
intelligence requires the ability to store and manipulate symbols
laws of qualitative structure
cell doctrine in biology
plate tectonics in geology
germ theory of disease
doctrine of atomism
“physical”
obey laws of physics
not restricted to human systems
designation - then given the expression, the system can affect the object
interpretation - expression designates a process
Heuristic searching#
symbol systems solve problems with heuristic search
Heuristic Search Hypothesis - solutions are represented as symbol structures. A physical symbol system exercises its intelligence in problem solving by search–that is, by generating and progressively modifying symbol structures until it produces a solution structure
from { cite newell1976computer }
there are practical limitations on how fast computers can search
To state a problem is to designate
a test for a class of symbol structures (solutions of the problem)
a generator of symbol structures (potential solutions).
To solve a problem is to generate a structure, using (2), that satisfies the test of (1).
searching is generally in a tree-form
Knowledge representation#
What a representation must support (representing, inferring, inferring efficiently, acquiring new knowledge) and two classic views: logic, with a syntax and a semantics, and associationist structures such as semantic networks and frames. Statistical, distributed representations come last.
physical symbol system hypothesis - a physical symbol system has the necessary and sufficient means for general intelligent action
computers and minds are both physical symbol systems
symbol - meaningful pattern that can be manipulated
symbol system - creates, modifies, destroys symbols
want to represent
meta-knowledge - knowledge about what we know
objects - facts
performance - knowledge about how to do things
events - actions
two levels
knowledge level - where facts are described
symbol level - lower
properties
representational adequacy - ability to represent
inferential adequacy
inferential efficiency
acquisitional efficiency - acquire new information
two views of knowledge
logic
a logic is a language with concrete rules
syntax - rules for constructing legal logic
semantics - how we interpret / read
assigns a meaning
multi-valued logic - not just booleans
higher-order logic - functions / predicates are also objects
multi-valued logics - more than 2 truth values
fuzzy logic - uses probabilities rather than booleans
match-resolve-act cycle
associationist
knowledge based on observation
semantic networks - objects and relationships between them - like is a, can, has
graphical representation
equivalent to logical statements
ex. nlp - conceptual dependency theory - sentences with same meaning have same graphs
frame representations - semantic networks where nodes have structure
ex. each frame has age, height, weight, …
when agent faces new situation - slots can be filled in, may trigger actions / retrieval of other frames
inheritance of properties between frames
frames can contain relationships and procedures to carry out after various slots filled
statistical
distributed - usually different from sparse code (sparser generally less robust)
opposite of sparse code = dense code
have to check multiple indexes
penti’s work: distributed
usually want these to be robust
nlp is main place where unsupervised pretraining widely used
hierarchical
good representations - linearly separable
representation that factors
information bottleneck method: want simple representation that keeps class but throws away lots of extraneous info
Expert systems#
An expert system stores a domain expert’s knowledge as condition-action rules over a database of facts and reasons by forward or backward chaining. MYCIN, an early success, used backward chaining and asked the patient questions.
expert system - program that contains some of the subject-specific knowledge of one or more human experts.
problems
planning
monitoring
instruction
control
need lots of knowledge to be intelligent
rule-based architecture - condition-action rules & database of facts
acquire new facts
from human operator
interacting with environment directly
forward chaining
until special HALT symbol in DB, keep following logical rule, add result to DB
conflict resolution - which rule to apply when many choices available
pattern matching - logic in the if statements
backward chaining - check if something is true
check database
check if on the right side of any facts
CLIPS - expert system shell
define rules and functions…
explanation subsystem - provide explanation of reasoning that led to conclusion
people
knowledge engineer - computer scientist who designs / implements ai
domain expert - has domain knowledge
knowledge engineering - art of designing and building expert systems
determine characteristics of problem
automatic knowledge-acquisition - set of techniques for gaining new knowledge
ex. parse Wikipedia
crowdsourcing
creating an expert system can be very hard
only useful when expert isn’t available, problem uses symbolic reasoning, problem is well-structured
MYCIN - one of first successful expert systems { cite shortliffe2012computer }
Stanford in 1970s
used backward chaining but would ask patient questions - sometimes too many questions
advantages
can explain reasoning
can free up human experts to deal with rare problems
Godel, Escher, Bach#
Hofstadter’s book is about self-reference: strange loops, Gödel’s incompleteness theorem, and how meaning arises when a formal system is interpreted. The AI notes list the abilities he thought intelligence needs and his picture of rules at several levels.
Douglas Hofstadter, 1979
meta
strange loop = paradox - self-referential
zen enlightenment
goal: transcend dualism = division into concepts (perception, words do this)
words give you some truth but always fail to describe some parts of the truth
music
canons - repeat w/ subtle changes (e.g. pitch shift)
fugue - repeat w/ more substantial changes
ai
essential abilities - can we do these things unsupervised?
to recognize the relative importance of different elements of a situation
to find similarities between situations despite differences which may separate them;
to draw distinctions between situations despite similarities which may link them
to synthesize new concepts by taking old concepts and putting them together in new ways
intelligence consists of rules at different levels
“just plain” rules - like reflexes which respond to stereotyped situations
metarules - when situations are mixtures of steretoyped situations, requires rules for deciding which “just plain” rules to apply
rules for inventing new rules - when situations can’t be classified
rules may have to change themselves
messages - comparison of DNA to a jukebox
where is info stored? records? buttons? smashed buttons?
what could constitute a Rosetta stone for DNA codes?
3 parts
frame - tells you that this is a message
outer - tells you how to read a message (e.g. language, style)
inner - actual content
if decoding is universal, we might call the outer message (e.g. the trigger) the message
memory - same bits can be used for different things - part of each message specifies the instruction type
brain
intelligence involves a calculus of descriptions = symbols
symbols represent both classes + instances (maybe both depending on amount of activation / context) - def need some context
can have links to other symbols (+priors on these)
top-down logical structure??
different ways to combine symbols get blurry
symbols can be learned to branch, merge
can harness temporal firing rates to encode more
can grow incrementally (greedily)
analogy of thoughts as trips on a (poorly fleshed out) map
interpretation
ex. top is decimal expansion of the sum of the second (\(\pi/4\))
7, 8, 5, 3, 9, 8, 1, 6, …
1, -1/3, +1/5, -1/7, +1/9, -1/11…
math / logic
godel’s thm - limitation of any formal axiomatic system: cannot make a program to find a complete + consistent set of axioms
church-turing thesis - a function on the natural numbers can be calculated by an effective method if and only if it is computable by a Turing machine
no system can do computation which cannot be broken down into simple elements
decision procedure - decides whether something is a theorem - must terminate
we can think of theorems as strings in a formal system
interpretation - correspondence between symbols and words
ideally, these are meaningful isomorphisms between codes and reality
not all interpretations imply meaningful (or valid) corresponding codes
there might be multiple, equally valid interpretations
consistency depends on interpretation:
consistency - when every theorem, upon interpretation, comes out true (in some imaginable world)
completeness - when all statements which are true (in some imageinable world), and which can be expressed as well-formed strings of the system, are theorems
slightly different axioms lead to elliptical/hyperbolic geometry instead of Euclidean geometry
godel numbering - can replace all symbols w/ numbers and all typographic rules w/ arithmetic rules
2 key idesas
strings can speak about other strings
self-scrutiny can be entire concentrated into a single string
every aspect of thinking can be viewed as a high-level description of a system which, on a low level, is governed by simple, even formal rules
causality
what counterfactuals are the most realistic
different things are stable at different levels
biology
dna -> rna -> proteins = sequence of amino acids
folds w/ valrious levels of structure (like music)
self-rep - what counts?
quine? instructions on jukebox? human reproduction?
Lesswrong#
A cognitive bias is a systematic error in judgment, like the bias of an estimator. Base rate neglect is the easiest to see with numbers.
-
cognitive bias can be thought of like statistical bias
base rate neglect: grounding one’s judgments in how well sets of characteristics feel like they fit together, and neglecting how common each characteristic is in the population at large
e.g. if you meet a shy person, are they more likely to be a salesperson or a librarian?
sunk cost fallacy
scope neglect - the number of birds saved—the scope of the altruistic action—had little effect on willingness to pay (post)
availability heuristic - judging the frequency or probability of an event by the ease with which examples of the event come to mind
absurdity bias; events that have never happened are not recalled, and hence deemed to have probability zero.
conjunction fallacy - humans assign a higher probability to a proposition of the form “A and B” than to one of the propositions “A” or “B” in isolation
the implausibility of one claim is compensated by the plausibility of the other; they “average out.”
planning fallacy - people think they can plan e.g. “best guess” scenarios are same as “best case” scenarios
rationality
epistemic rationality - systematically improving the accuracy of your beliefs
instrumental rationality - systematically achieving your values
System 1 and System 2—fast perceptual judgments versus slow deliberative judgments. System 2’s deliberative judgments aren’t always true, and System 1’s perceptual judgments aren’t always false; so it is very important to distinguish that dichotomy from “rationality.”
Key ideas
Nativism, empiricism and constructivism differ in how much of a representation is built in and how much is learned from experience.
Bayesian models of cognition work at Marr’s computational level and ask which priors would make people’s inferences rational.
Deep learning takes the empiricist side, and in these notes less built-in structure plus more data has usually worked better.
Chomsky’s argument against behaviorism is that children say things they have never heard, so reinforcement alone cannot explain grammar.
Classic AI represents knowledge as symbols manipulated by search, as logic, or as semantic networks and frames, and expert systems apply rules by forward or backward chaining.
A cognitive bias is a systematic error in judgment, like the bias of an estimator; base rate neglect ignores how common each category is.