Transcript Document

Reopening the Critical Period:
Understanding and Remediating
Failures of Learning
James L. McClelland
Center for the Neural Basis of Cognition
CNBC
A Joint Project of Carnegie Mellon and the University of Pittsburgh
The Traditional Approach to Modeling
Human Cognition
• The symbolic information processing framework:
– The mind is viewed as a type of computer
– Knowledge is coded explicitly in propositions
(structured collections of symbols) and rules that
manipulate them.
– Learning occurs via adding, refining, or
strengthening the rules and propositions.
– There is little contact with the our growing
understanding of the neural mechanisms
underlying cognition in the brain.
An Alternative Approach that Bridges to
Neuroscience
• Awakening from the Cartesian dream:
– Constructing theories of mental processes that
exploit the mechanisms that are found in the
brain.
– Capitalizing on the use of evidence from
neuroscience as well as behavior to guide theory
and experiment.
What is the Neural Basis of Mental
Processes?
• Mental processes occur via
the propagation of excitatory
and inhibitory signals among
neurons via weighted
synaptic connections.
Why Adopt this View?
• It has helped us to construct explicit mechanistic models that
– Overcome the rigidity of conventional ‘symbolic’ models of
human perception, cognition, and action.
– Capture the flexibility and fluidity of human mental abilities.
– Provide detailed accounts of the patterns of findings
obtained in behavioral experiments with human subjects.
– Revisit questions about the nature of mental processes and
representations.
– Develop new ways of thinking about learning that reopen old
questions about innateness, learnability, and critical periods.
What is the Knowledge Underlying
Mental Processes?
• If mental processes occur via
the propagation of excitatory
and inhibitory signals among
neurons via weighted
synaptic connections.
• Then the knowledge that
guides these processes is
stored in the strengths of the
connections among the
neurons.
If this is true, then what about learning?
• If the knowledge underlying
cognition is stored in the
strengths of the connections
among the neurons.
• Then
– Learning must occur
through the adjustment
of the strengths of the
connections.
Hebb’s Postulate
“When an axon of cell A is near enough to excite a
call B and repeatedly or persistently takes part in
firing it, some growth process or metabolic change
takes place in one or both cells such that A’s
efficiency, as one of the cells firing B, is increased.”
D. O. Hebb, Organization of Behavior, 1949
In other words:
“Cells that fire together wire together.”
Unknown
An Experiment Showing That LTP Is
“Hebbian”
Glutamate ejected from
the pre-synaptic terminal
activates AMPA
receptors, exciting the
post-synaptic neuron.
Glutamate binds to the
NMDA receptor but only
becomes active when the
neuron is excited.
Ca++ flows in, increasing
AMPA receptors.
Studies of Cortical Reorganization in
Monkeys
• Alteration of experience leads to alterations of neural
representations in the brain.
• What neurons represent, and how precisely they
represent it, are strongly affected by experience.
• We allocate more of our brain to things we have the
most experience with.
• The effects appear to be consistent with Hebbian,
competitive models of learning.
Monkey Somatosensory Cortex
Merzenich’s Joined Finger Experiment
Merzenich’s Rotating Disk Experiment
Merzenich’s Rotating Disk Experiment:
Redistribution and Shrinkage of Fields
Merzenich’s Rotating Disk Experiment:
Expansion of Sensory Representation
Plusses and Minuses of Hebbian Learning
• Hebbian learning tends to reinforce whatever
response occurs to a particular input.
• This may contribute to failures and pathologies of
learning and even to stamping in of bad habits when
the response we make to an input is not the best
one to make.
• Possible examples:
- Dystonia in musicians and writers
- Phobias, racism
- Entrenchment of “Habits of Mind”
- Problems with new phonetic distinctions in
second language learners.
Why Can’t Japanese Adults Learn to
Distinguish “r” and “l”?
• In their native language they hear “r” and “l” as the
same.
• When they hear “r” or “l”, Hebbian learning may
unhelpfully reinforce this tendency.
• A simple neural network model based on
competitive, Hebbian learning illustrates how this
can happen.
Network Architecture and Function
• Kohonen network with initial
weak topographic biases
• Input projects to
representation layer where
unit with strongest input is
selected
• Winner and near neighbors
are activated with Gaussian
falloff
• Learning occurs according to
the Hebb-like Oja rule:
Dwrs = ear(as - wrs)
Training Environment
Background
“phoneme”
inputs
“English” /r/
and /l/ inputs
“Japanese”
/r/-like input
Learning in Standard Environment Only
Vs. Foreign Then Standard Environment
English Only
Japanese then English
Implications for Teaching /r/-/l/
Discrimination
• If the Hebbian explanation is correct, perhaps we
can help Japanese adults learn by using exaggerated
inputs that they can distinguish.
• We can illustrate this in a continuation of the
simulation, using exaggerated versions of the /r/ and
/l/-like inputs:
Standard Inputs
Exaggerated Inputs
Breaking up the Single Percept Using
Exaggerated Inputs
Japanese
English
E+E
English
Our Experiment
• We wanted to test the idea that exaggerated inputs
would lead to rapid learning.
• We used an adaptive training procedure, starting
with exaggerated inputs.
• According to the Hebbian hypothesis, no feedback is
necessary to learn, so we presented stimuli without
feedback.
• In the fixed training condition, we used stimuli that
were very hard for our subjects to discriminate
initially.
A Continuum of Speech Sounds
“lock”
“rock”
Fixed and Exaggerated Stimuli
• Fixed stimuli
– “lock”
– “rock”
• Starting exaggerated
stimuli
Fixed
Exaggerated
rock
lock
lock
rock
– “lock”
rock
– “rock”
lock
Experimental Details
• Eight subjects received fixed training, eight subjects
received adaptive training.
• Only subjects performing below 70% correct in
pretest identification of fixed training stimuli were
included in either group.
• Half of subjects received training with “rock”-”lock”;
the other half received training with “road”-”load”.
• One each trial, the “r” or the “l” stimulus was
presented. Subjects responded by pressing a key to
choose “r” or “l”. No feedback was given.
• Training took place over three, 20-minute sessions,
each involving 480 training trials and 50 probe trials.
Adaptive and Fixed Training
• In the adaptive condition:
– Stimuli were adjusted inward (more difficult) after
8 correct responses.
– Stimuli were adjusted outward (more easy) after
every error.
– Two trials in every 20 were probe trials using the
fixed training stimuli.
• In the fixed condition:
– The same fixed stimuli were used throughout
training.
Time Course
of Training in
Three Subjects
Two Further Questions
• Does feedback play any role?
– No effect of feedback is predicted from the
Hebbian theory.
– To examine this, we added two more conditions:
> Fixed Training with Feedback
> Adaptive Training with Feedback
• What are the effects of additional training?
– Half of subjects in each group trained for three
additional days
Summary
• Adaptive training is successful, with or without feedback.
• Feedback makes a big difference for learning with difficult, fixed
stimuli.
• Transfer does occur when training is successful, with some lag
and some decrement.
• Six days of training produces robust learning with transfer in all
conditions except the fixed no feedback condition.
• There is some evidence of learning even in the fixed-no
feedback condition. However, this learning may be coming
from exposure to exaggerated stimuli during testing sessions.
A Further Issue and a Follow-up
Experiment
• Why is it possible to make such rapid progress in this
experiment?
– Based on our model, we suggest that allowing subjects to
focus on /r/ vs. /l/ in a single context greatly facilitates
learning.
– In the network, we find that learning will occur fairly quickly
if we simply eliminate the background stimuli.
– The ‘representational space’ is partitioned among the
training stimuli; when we have only /r/ vs. /l/ they soon ‘spit
apart’, dividing the space.
• To test this, we plan to compare training on /r/ vs. /l/ with and
without other stimuli mixed in with the training items
(e.g., /m/, /n/, /y/, /h/, etc.)
Future Directions for the Theory of
Learning
• It appears that the “Hebbian theory” is incomplete, since it
provides no role for feedback.
• We are exploring two biologically motivated alternatives:
– Error-modulated Hebbian learning:
• Increases learning rate when feedback indicates difficulty.
– Reward-modulated Hebbian learning
• Learning is enhanced when an unexpected “reward” occurs.
• In these models, learning can occur in the absence of feedback.
This is important to account for passive exposure learning and
learning without feedback in our experiments.
• In both models, feedback enhances learning. The modulation
of Hebbian learning ‘reweighs the competition for
representational space, giving stimuli the subject is having
difficulty with more room in the representation.
Future Directions for the Framework
• As in the case of the /r/-/l/ work, we seek to continue adapting
the framework to incorporate principles from neuroscience.
• The goal, as in the case of Hebbian learning, is to consider how
these principles impact on our understanding of human mental
functions, such as perceptual learning.
• The long-term goal is a theoretical framework that will have
explicit links to neuroscience, and encompass
– Normal cognitive and other processes
– Disorders of mental processes
– Development and remediation