Learning, development and plasticity

Josh TenenbaumMIT

Department of Brain and Cognitive SciencesComputer Science and AI Lab (CSAIL)

Can we develop an integrated theory of learning, in computational, behavioral and neural terms?

• The “standard model”

t

n

ttt

n

ttt

xxwywEw

xwyE

1

1

2

)(

)(21

t

n

tt

n

t

n

ttt

xywEw

xwyE

1

1 1

22 )(21

21

Sup

ervi

sed

Uns

uper

vise

d

amF

The really hard problems

The two cultures

“Nurture”Statistical learning, plasticityGrounded in the brain

“Nature”Innate structured representationsGrounded in cognition

versus

Recent causes for optimism• New models from machine learning, AI

– Structured statistical modelsProbabilities defined over structured representations: graphs,

causal networks, grammars, predicate logic.

– Multilevel (hierarchical) statistical modelsInference at multiple levels of abstraction and multiple

timescales.

– Flexible statistical modelsHypothesis spaces grow as new data are encountered.

• New technologies – “Supercomputers” on the desktop, grid computing– Life-size datasets for modeling cognitive development– Mainstream functional MRI

Phrase structure

Utterance

Speech signal

Grammar

“Universal Grammar” Hierarchical phrase structure grammars (e.g., CFG, HPSG, TAG)

P(phrase structure | grammar)

P(utterance | phrase structure)

P(speech | utterance)

P(grammar | UG)

Left IFG

VerbVPNPVPVP

VNPRelRelClauseRelClauseNounAdjDetNP

VPNPS

][][][

Probabilistic scene parsing

1 2

1 2

t=0

t=1

t=2

Dest=1

Dest=1

Dest=2

Dest=2

BLOG(Bayesian Logic)

Probabilistic scene parsing

Learning domain structures

F: form

S: structure

D: data

tree

mammal

Tree with objects at leaf nodes

plant

animal

living thing

P(structure | form)

P(data | structure)

mouse

gorilla

pine

palm

Learning domain structures

AbstractPrinciples

Structure

Data

Learning causal theories

Learning regulatorymodules in genetics

Learning the lawsof magnetism

Causal learning with complementary “cortical” and

“hippocampal” networks

Learning to act

Concepts learned:Clear, InhandTopstackAboveHeight…

Exploration vs. Exploitation

Constraints Goals

Actions

Rational planning(PO)MDP

Understanding actions:goal inference

Model Predictions

Hum

an

judg

men

ts

The grand challenge

Cognitive science of human learning

Design of artificial learning systems

Brain structures and mechanisms that support learning