McClelland226IntroCompLearnSys - Psychology

advertisement
Cooperation of Complementary Learning
Systems in Memory
Review and Update on the Complementary
Learning Systems Framework
James L. McClelland
Psychology 226
Fall, 2008
Stanford University
A Playwright’s Take on Memory
“What interests me a great deal is the
mistiness of the past”
Harold Pinter, Conversation prior to
the opening of Old Times, 1971
What is a Memory?
• The trace left by an experience.
• A representation of the experience brought
“back to mind” later.
• In some theories, these things are one and
the same
• Not so in a connectionist approach to memory!
In a connectionist approach…
•
An experience produces a pattern
of activation over an ensemble of
processing units.
•
The memory trace is a pattern of
adjustments to connections among
simple processing units.
•
The memory as recalled is a
pattern of activation constructed
with the help of the affected
connections.
•
Connections are affected by many
experiences, so ‘recall’ is always
subject to influence from traces of
other experiences.
•
Remembering is thus always a
process of reconstruction.
Outline
 What is “a memory”?
 The essence of the connectionist/PDP perspective
• Contrasting systems-level approaches to the neural
basis of memory
• The Complementary Learning Systems framework
• McClelland, McNaughton, and O’Reilly, 1995
• How the complementary learning systems work
together to create ‘episodic’ and ‘semantic’ memory.
Outline
 What is “a memory”?
 The essence of the connectionist/PDP perspective
 Contrasting systems-level approaches to the neural
basis of memory
• The complementary learning systems approach
• McClelland, McNaughton, and O’Reilly, 1995
• How the complementary learning systems work
together to create ‘episodic’ and ‘semantic’ memory.
Multiple Memory Systems
• Seeks dissociations of different forms of learning and
memory.
–
–
–
–
Explicit vs. implicit memory
Declarative vs. procedural memory
Semantic vs. episodic memory
Familiarity vs. recollection
• Although more than one system can contribute to
performance in a given task, the contributions are
simply alternative paths to correct performance.
• E.g., in a recognition memory task:
– One can respond either by familiarity or recollection:
p(old) = p(recall) + (1-p(recall)) * p(familiar)
An Alternative Approach
• Complementary and Cooperating Brain
Systems
– Memory task performance depends on multiple
contributing brain systems.
– Contributions of each system to overall task
performance depend on their neuro-mechanistic
properties.
– Systems work together so that overall performance
may be better than the sum of the independent
contributions of the parts.
The Complementary Learning Systems
Theory
(McClelland, McNaughton & O’Reilly, 1995)
• Neuropsychological motivation
• The basic theory
• Neurophysiology consistent with the account
• Why there should be complementary systems
Bi-lateral destruction
of hippocampus and
related areas
produces:
- Profound deficit in forming
new arbitrary associations
and new episodic
memories.
- Preserved acquisition of
skills and item-specific
priming.
- Loss of recently learned
material w/ preservation
of prior knowledge,
acquired skills, and
remote memory.
Control groups
Lesioned groups
Time from experience to lesion in days
The Neuro-Mechanistic Theory:
Processing and Learning in Neocortex
• An input and a response to
it result in activation
distributed across many
areas in the neocortex.
• Small connection weight
changes occur as a result,
producing
– Item-specific effects
– Gradual skill acquisition
• These small changes are
not sufficient to support
rapid acquisition of
arbitrary new associations.
Complementary Learning System in the
Hippocampus
• Bi-directional connections
produce a reduced
description of the cortical
pattern in the hippocampus.
• Large connection weight
changes bind bits of reduced
description together
• Cued recall depends on
pattern completion within the
hippocampal network
• Consolidation occurs through
repeated reactivation,
leading to cumulation of
small changes in cortex.
hippocampus
Supporting Neurophysiological
Evidence
• The necessary pathways
exist.
• Anatomy and physiology of
the hippocampus support its
role in fast learning.
• Reactivation of hippocampal
representations during
sleep.
Different Learning and Coding
Characteristics of Hippocampus and
Neocortex
• Hippocampus learns quickly to allow one-trial learning
of particulars of individual items and events.
• Cortex learns slowly to allow sensitivity to overall
statistical structure of experience.
• Hippocampus uses sparse conjunctive representations
to maintain the distinctness of specific items and
events.
• Cortex uses representations that start out highly
overlapping and differentiate gradually to allow:
– Generalization where warranted
– Differentiation where necessary
Examples of neurons found in entorhinal cortex and
hippocampal area CA3, consistent with the idea that the
hippocampus but not cortex uses sparse conjunctive
coding
Recording was made while animal traversed an eight-arm radial maze.
Why Are There Complementary Learning
Systems?
• Discovery of structure
requires gradual interleaved
learning with dense
(overlapping) patterns of
activation. (Many aspects
of semantic cognition and
conceptual development are
explained by this approach).
• Rapid learning of new
information in such systems
leads to catastrophic
interference.
• The hippocampus (working
with the cortex) can solve
this problem.
Keil, 1979
The Model of Rumelhart (1990)
Differentiation in Development
Initially
Still Young
Somewhat Older
Catastrophic
Interference
• First observed by McClosky
and Cohen (1989) when
they tried to teach first one,
then another list to a neural
network.
• All items on the first list
were forgotten before even
one item from the second
list was learned.
• Catastrophic interference
also occurs if one tries to
teach the trained Rumelhart
network some partially
inconsistent new
information.
How can we solve
the problem?
• Hippocampus provides a
separate system that can
learn the new information
rapidly.
• Once in the hippocampus,
the information can be
reinstated, allowing cortical
learning.
• If these hippocampal
reinstatements are
interleaved with ongoing
exposure to other items,
the new information will be
integrated into the cortical
system without interfering
with what is already known.
How can we solve
the problem?
• Hippocampus provides a
separate system that can
learn the new information
rapidly.
• Once in the hippocampus,
the information can be
reinstated, allowing cortical
learning.
• If these hippocampal
reinstatements are
interleaved with ongoing
exposure to other items,
the new information will be
integrated into the cortical
system without interfering
with what is already known.
How can we solve
the problem?
• Hippocampus provides a
separate system that can
learn the new information
rapidly.
• Once in the hippocampus,
the information can be
reinstated, allowing cortical
learning.
• If these hippocampal
reinstatements are
interleaved with ongoing
exposure to other items,
the new information will be
integrated into the cortical
system without interfering
with what is already known.
Overview
 What is “a memory”?
• The essence of the connectionist/PDP perspective
 Contrasting systems-level approaches to the neural
basis of memory
 The complementary learning systems approach
• McClelland, McNaughton, and O’Reilly, 1995
 How the complementary learning systems work
together to create ‘episodic’ and ‘semantic’ memory.
Kwok & McClelland Model of
Semantic and Episodic Memory
• A slow learning cortical system
and a fast-learning hippocampal
system.
• Cortex contains units
representing both content and
context of an experience.
• “Semantic” memory is gradually
built up through repeated
presentations of the same
content in different contexts.
• Memory for a specific episode
depends on hippocampus and the
relevant cortical areas, including
context.
• Episodic memories benefit from
relevant semantic learning.
• Virtually all memories are partly
semantic and partly episodic.
Hippocampus
Context
Neo-Cortex
Relation
Cue
Target
Effect of Prior Association on PairedAssociate Learning in Control and
Amnesic Populations
Cutting (1978), Expt. 1
100
Control (Expt)
Percent Correct
80
Amnesic (Expt)
60
40
20
0
Base rates
-20
Very Easy
Easy
Fairly Easy
Hard
Category (Ease of Association)
Very Hard
Kwok & McClelland Simulation:
Cortical Pre-Training
•
Cortical network is pre-trained
using CHL with 4 cue-relationtarget triples for each of 20
different cues.
– Dog chews bone
– Dog chases cat
– …
•
Context varies randomly
throughout cortical pretraining.
•
Words and context are patterns
of activation over units in the
appropriate pool.
•
Training frequency was varied
to create strong and weak
associates for each cue.
Hippocampus
Context
Neo-Cortex
Relation
Cue
Target
Kwok & McClelland Simulation:
Experiment
• Study phase is simulated by
presenting a set of cue-target
pairs in a fixed context.
• Cortex fills in relation as
mediator.
• Hippocampal network assigns
sparse conjunctive
representation to the
combined cue and context.
• Hebbian learning is used to
associate this representation
with the corresponding target
pattern.
• At test, context and cue are
presented, cortex and
hippocampus collaborate to
fill in target.
• Amnesia is simulated by
removing some (or all) of the
hippocampal units.
Hippocampus
Context
Neo-Cortex
Relation
Cue
Target
Simulation Results From KM Model
Cutting (1978), Expt. 1
100
80
Percent Correct
Control (Model)
84
Amnesic (Model)
70
68
Control (Expt)
60
Amnesic (Expt)
40
20
9
0
0
0
-20
Very Easy
Easy
Fairly Easy
Hard
Category (Ease of Association)
Very Hard
Summary
• Memory traces are in your connections;
memories are constructed using these traces
(and those of other experiences) to constrain
the construction process.
• Memory task performance involves
cooperation among brain regions:
– Cortical regions that gradually learn to represent
content and context
– Medial temporal regions that can learn conjunctive
associations of cortical patterns rapidly
• There are no separate systems dedicated to
different kinds of memory. These functions
depend on cooperating brain systems.
Download