Cognitive processing as constraint satisfaction

43
Cognitive processing as constraint satisfaction

description

Cognitive processing as constraint satisfaction. Today’s goals. One example of how “neutrally-inspired” mechanisms can address problems faced by symbolic approaches (in memory) One example of how models can be used formally as proofs of general principles. Introduction to HW1 models. - PowerPoint PPT Presentation

Transcript of Cognitive processing as constraint satisfaction

Page 1: Cognitive processing as constraint satisfaction

Cognitive processing as constraint satisfaction

Page 2: Cognitive processing as constraint satisfaction

Today’s goals

• One example of how “neutrally-inspired” mechanisms can address problems faced by symbolic approaches (in memory)

• One example of how models can be used formally as proofs of general principles.

• Introduction to HW1 models.

Page 3: Cognitive processing as constraint satisfaction

Aristotle: Memory is like a library.

Everyone else for next 2300 years: Yup.

Cognitive neuroscience: LOL, wut?

Page 4: Cognitive processing as constraint satisfaction

How like a library?

Page 5: Cognitive processing as constraint satisfaction

Ways memory is not like a library

• Content-addressable

• Generalizable

• Graceful degradation

• Graded, uncertain

• Supported by neural systems

Page 6: Cognitive processing as constraint satisfaction
Page 7: Cognitive processing as constraint satisfaction
Page 8: Cognitive processing as constraint satisfaction
Page 9: Cognitive processing as constraint satisfaction
Page 10: Cognitive processing as constraint satisfaction

Demo

Page 11: Cognitive processing as constraint satisfaction

Ways memory is not like a library

• Content-addressable

• Generalizable

• Graceful degradation

• Graded, uncertain

• Supported by neural systems

Page 12: Cognitive processing as constraint satisfaction

It does something, but how do you know it is doing the “right” thing?

Page 13: Cognitive processing as constraint satisfaction

• Suppose all weights are symmetrical• Weights represent “beliefs” about the

mutual compatibility of two features• If w > 0, “If A is on B should be on

and vice versa”• If w = 0, “A has no implications for B• If w < 0, “If A is on B should be off

and vice versa”• So for any pair of features you can

measure how well their activations “fit” the belief about their compatibility as:

G_ij = a_i * a_j * w_ij

G = How “good” is the pattern of activation, given knowledge of the mutual constraints between features?

Page 14: Cognitive processing as constraint satisfaction

To get a measure of how well the pattern across the whole network satisfies the “beliefs” coded in the weights you can sum this over all pairs of units…

Patterns that “fit” the constraints built into the weights and the external input will produce a large value in G…

An activation function that increases toward its maximum when net input is positive and decreases it toward zero when net input is negative is guaranteed to increase G for the whole network!

In other words, a local computation (should I raise or lower my behavior?) leads to system-wide improvement in “fitting” the constraints.

Page 15: Cognitive processing as constraint satisfaction
Page 16: Cognitive processing as constraint satisfaction
Page 17: Cognitive processing as constraint satisfaction

Schemata as constraint satisfaction

What do you know about restaurants?

Page 18: Cognitive processing as constraint satisfaction

• You eat there• May be a waiter• Order food• Sit down• Obtain food• Eat food• Pay for food• Leave tip

Page 19: Cognitive processing as constraint satisfaction

• Order food• At counter• From waiter

• Sit down• Choose seat• Shown to seat

• At table• At counter

• Obtain food• At counter• From waiter

• Eat food

• Pay for food• Before getting it• After getting it

• Use cash• Use credit

• Leave tip

Variables

Values

SchemaContains variables

Codes relations between variables

Contains constants

Variables

Specify different possible “fillers”

Might specify default values

Page 20: Cognitive processing as constraint satisfaction

Desiderata

• Should capture structure– Information about common elements,

variables, distributions of values for variables

• Should capture sub-structure– The “slot” in a given schema might be filled by

another more specific schema– E.G. “Paying”

• Should be flexible and applicable to new situations.

Page 21: Cognitive processing as constraint satisfaction

Problems• Not actually very flexible

– Eating in a dark restaurant?– Seems like you need to store all possible structures, implausible.– How can structured representations be adapted “on the fly”?

• Overly structured– Values of “fillers” for some variables depend on how others have been filled.– Not clear how to combine different structured representations– Relational information in schema depends on values of variables contained

• E.g. Order information of “Sitting” versus “Ordering” depends on value of variable “Obtain food”

• Acquisition– How do you learn these structures in the first place?

Page 22: Cognitive processing as constraint satisfaction

Similar issues arise in:

• Grammar

• Event representation

• Spatial representation

• Transitive inference

• Logic / reasoning

• Analogy

Page 23: Cognitive processing as constraint satisfaction

Schematic knowledge as constraint satisfaction

Page 24: Cognitive processing as constraint satisfaction

CeilingWallsDoor

WindowVery-large

LargeMedium

SmallVery-small

DeskTelephone

BedTypewriterBookshelf

CarpetBooks

Desk-chairClock

PictureFloor-lamp

SofaEasy-chairCoffee-cup

AshtrayFireplace

DrapesStove

SinkRefrigerator

ToasterCupboardCoffeepot

DresserTV

BathtubToiletScale

CoathangerComputer

Oven

Kitchen Office Livingroom Bedroom Bathroom

Page 25: Cognitive processing as constraint satisfaction
Page 26: Cognitive processing as constraint satisfaction

CeilingWallsDoor

WindowVery-large

LargeMedium

SmallVery-small

DeskTelephone

BedTypewriterBookshelf

CarpetBooks

Desk-chairClock

PictureFloor-lamp

SofaEasy-chairCoffee-cup

AshtrayFireplace

DrapesStove

SinkRefrigerator

ToasterCupboardCoffeepot

DresserTV

BathtubToiletScale

Cei

ling

Wal

lsD

oor

Win

dow

Ver

y-la

rge

Larg

eM

ediu

mS

mal

lV

ery-

smal

lD

esk

Tel

epho

neB

edT

ypew

riter

Boo

kshe

lfC

arpe

tB

ooks

Des

k-ch

air

Clo

ckP

ictu

reF

loor

-lam

pS

ofa

Eas

y-ch

air

Cof

fee-

cup

Ash

tray

Fire

plac

eD

rape

sS

tove

Sin

kR

efrig

erat

orT

oast

erC

upbo

ard

Cof

feep

otD

ress

erT

VB

atht

ubT

oile

tS

cale

Coa

than

ger

Com

pute

rO

ven

Page 27: Cognitive processing as constraint satisfaction
Page 28: Cognitive processing as constraint satisfaction

“Variables” coded in weight structure…

Page 29: Cognitive processing as constraint satisfaction
Page 30: Cognitive processing as constraint satisfaction

Function (e.g., making inferences) achieved by activation updating / hill-climbing…

Page 31: Cognitive processing as constraint satisfaction
Page 32: Cognitive processing as constraint satisfaction

Different “schemas” correspond to different localmaxima in G….

Page 33: Cognitive processing as constraint satisfaction
Page 34: Cognitive processing as constraint satisfaction

Goodness landscape determines how discrete /combinable different schemas are…

Page 35: Cognitive processing as constraint satisfaction
Page 36: Cognitive processing as constraint satisfaction

“Schemas” adapt themselves to the situation specifiedby the input…

Page 37: Cognitive processing as constraint satisfaction
Page 38: Cognitive processing as constraint satisfaction
Page 39: Cognitive processing as constraint satisfaction

“Coalitions” of connected units can form “sub-schemas”to provide for different levels of structure…

Page 40: Cognitive processing as constraint satisfaction
Page 41: Cognitive processing as constraint satisfaction

Homework 1 begins this week!

Page 42: Cognitive processing as constraint satisfaction

20s 30s 40s Pusher Burglar Bookie

Jet Shark JH HS Col.

Page 43: Cognitive processing as constraint satisfaction

20s 30s 40s Pusher Burglar Bookie

Jet Shark JH HS Col.

Art-in Ken-in