Modeling Computational of Human Cognition
Professor Michael C. Mozer CSCI 3202
Computational Modeling
Computer simulation of neural and/or cognitive processes that underlie performance on a task Goals
Understand mechanisms of information processing in the brain Explain behavioral, neuropsychological, and neuroscientific data Suggest techniques for remediation of cognitive deficits due to brain injury and developmental disorders Suggest techniques for facilitating learning in normal cognition Construct computer architectures to mimic human-like intelligence
Why Build Models?
Forces you to be explicit about hypotheses and assumptions Provides a framework for integrating knowledge from various fields Allows you to observe complex interactions among hypotheses Provides ultimate in controlled experiment Leads to empirical predictions A mechanistic framework will ultimately be required to provide a unified theory of cortex.
Levels of Modeling
Single cell
ion flow, membrane depolarization, neurotransmitter release, action potentials, neuromodulatory interactions
Network
neurophysiology and neuroanatomy of cortical regions, cell firing patterns, inhibitory interactions, mechanisms of learning
Functional
operation and interaction of cortical areas, transformation of representations
Computational
input-output behavior, mathematical characterization of computation
Overview
Computational modeling Modeling human learning Modeling performance after brain damage
Using Testing to Enhance Learning: A Comparison of Two Hypotheses
Michael C. Mozer Michael Howe
Department of Computer Science and Institute of Cognitive Science University of Colorado, Boulder
Harold Pashler
Department of Psychology University of California, San Diego
Fact Learning
E.g., foreign language vocabulary
French word for dog is chien.
E.g., history trivia
The University of Colorado was founded in 1876.
Facts can be framed as cue response pairs.
e.g., dog chien e.g., Founding of University of Colorado 1876
a.k.a. paired associate learning
Self Testing
dog table house
Does Self Testing Foster Learning?
Long history of empirical demonstrations, but many methodological difficulties.
Carrier and Pashler (1992)
Pure study (PS)
cue-response pair presented together for 10 sec. cue response 0 5 10 sec. cue response
Self testing (ST)
cue presented alone for 5 sec., during which response is to be retrieved cue and response together for 5 sec., during which response is to be studied
Design
20 items designated for PS, and 20 for ST 3 training blocks; all items studied in block 1 final test phase; evaluation via cued recall Experiment 1: consonant trigrams two-digit numbers Experiment 2: English Siberian Eskimo Yupik word translation
Carrier and Pashler (1992)
40
% Error
30 20 10 PS ST PS ST
Expt. 1
Expt. 2
Carrier and Pashler (1992)
40
% Error
30 20 10 PS ST PS ST
Expt. 1
Possible explanation
Expt. 2
Subjects used first self-test trial to assess item difficulty, and increased encoding effort on second self-test trial.
Expt. 3 same as Expt 2. except all items studied in first two training blocks
Carrier and Pashler (1992)
40
% Error
30 20 10 PS ST PS ST PS ST
Expt. 1
Expt. 2
Expt. 3
Some Explanations of Self-Testing Benefit
Landauer and Bjork (1978)
Retrieval attempts provide general boost to performance at a future time. Incorrectly predicts that ST and PS items should benefit equally
Mandler (1979)
Cued recall strengthens structural, integrative information about an item. Because cue and response are simultaneously activated for both ST and PS items, not clear why they wouldnt both benefit.
Bjork (1975)
Act of retrieval strengthens existing retrieval routes to the response. Consistent with data, but seems to require novel learning mechanisms
Basic Approach
Use a common, relatively noncontroversial architecture Feedforward neural network
Input layer connected to output layer Standard sigmoidal activation function Error correction learning
Widrow-Hoff (a.k.a. LMS) network generates actual output teacher provides target output training depends on actual target
target output output layer
input layer
Basic Approach
Use a common, relatively noncontroversial architecture Feedforward neural network
Input layer connected to output layer Standard sigmoidal activation function Error correction learning
Widrow-Hoff (a.k.a. LMS) network generates actual output teacher provides target output training depends on actual target
target output output layer
input layer
Training of neural net often viewed as abstract procedure for loading knowledge into net. Here, we make a stronger claim.
One training trial in neural net ~ one experimental trial
Hypothesis 1: Self-Generated Training Targets
Guthrie (1952)
One learns what one does.
ST item
Self test Study candidate response correct response target for error-correction learning target for error-correction learning
Both candidate and correct response are trained.
PS item
Only correct response is trained.
Choosing candidate response
Probabilistic selection with Luce Choice Rule (a.k.a. Boltzmann distribution)
response 1 response 2
response 3 response 4
Hypothesis 1: Simulation Result
No parameter settings found that yield an enhancement of learning by testing. In final test, mean-squared error (MSE) significantly higher for ST than PS items.
MSEST MSTPS
0.6 0.4 0.2 0 1
1 le a rn 0.5 0.5 me ing r lfn t - p a te or se 0 0 ef rov for g rat target id e e x p in d ta eri learn erated r ge gen t
Hypothesis 2: Interruption of Cue Processing
Carrier and Pashler (1992)
Presentation of the response simultaneously with cue interrupts processing of the cue, making learning less efficient.
Hypothesis 2: Interruption of Cue Processing
Carrier and Pashler (1992)
Presentation of the response simultaneously with cue processing interrupts of the cue, making learning less efficient.
Our interpretation
Units in neural net have temporal dynamics. Leaky integrator model: y i ( t ) = y i ( t 1 ) + ( 1 )f ( net i ( t ) ) target asymptote
target output output layer
output unit activity
time
input layer
Hypothesis 2: Interruption of Cue Processing
Carrier and Pashler (1992)
Presentation of the response simultaneously with cue interrupts processing of the cue, making learning less efficient.
Our interpretation
Units in neural net have temporal dynamics. Leaky integrator model: y i ( t ) = y i ( t 1 ) + ( 1 )f ( net i ( t ) ) target asymptote actual time Presentation of correct response premature termination of processing incorrect output incorrect error signal
target output output layer
output unit activity
input layer
Hypothesis 2: Simulation Result
0.4
Squared Error
0.3 0.2 0.1 0 PS ST PS ST
Expt. 1 & 2
Expt. 3
Summary
Goal
Explain the enhancement of learning through testing
Approach
In the context of a simple neural network model, we explored two alternative hypotheses. (1) Testing obtains a candidate response whose association to the cue is strengthened, making the association less vulnerable to decay or interference. (2) Error-correction learning requires comparison of the correct response to a candidate response. Testing forces a candidate response to be generated, whereas pure study does not.
Result
Simulations supported hypothesis 2, not hypothesis 1.
Modeling Neuropsychological Phenomena
Michael C. Mozer Mark Sitton
Department of Computer Science and Institute of Cognitive Science University of Colorado, Boulder
Martha Farah
Department of Psychology University of Pennsylvania
Interactivity in Brain Damage
Neuropsychological disorders have traditionally been explained by a focal lesion to a single processing pathway.
Farah (1990) argued that certain highlyselective deficits might have a parsimonious account in terms of multiple lesions with interactive effects.
We illustrated the viability of this account via a neural network model of optic aphasia.
Optic Aphasia
Deficit in naming visually presented objects, in the absence of visual agnosia and general anomia
Nonverbal indications of recognition: sorting, gesturing Naming possible given verbal definition, tactile stimulation, object sounds
Visual system roughly intact
Insensitivity to visual quality; can copy drawings; normal interaction with world
No prosopagnosia Alexia Neuropathology: unilateral left posterior lesions, including occipital cortex and white matter Past accounts have postulated multiple semantics systems or multiple functional pathways to naming.
Cognitive Architecture
name semantic visual auditory gesture
Each arrow represents a processing pathway (neural net) Pathway act as associative memories
Proposed Explanations are Unparsimonious
name semantic visual auditory visual gesture name semantic auditory gesture
name semantic visual auditory visual semantics visual
name verbal semantics auditory
name left hemi semantics visual right hemi semantics auditory
gesture semantic visual
name
auditory
Neural Network Implementation of Pathway
pathway output
clean up: recurrent attractor network
mapping: multilayer feedforward network
pathway input
input space
mapping clean up
output space
Model Dynamics
attractor units a(t) e(t)
attractor unit update equation: a j ( t ) = exp ( s ( t ) j aj ( t ) = aj ( t ) ai ( t ) state unit update equation:...