 n
this interview, Dr. James McClelland, Dr. Bruce McNaughton, and
Dr. Randall O’Reilly discuss their highly cited paper,
"Why there are complementary learning systems in the
hippocampus and neocortex—insights from the successes and
failures of connectionist models of learning and memory,"
(Psychol. Rev. 102[3]: 419-57, July 1995). This paper
has been cited 477 times to date and currently ranks at #15
among Psychiatry/Psychology papers published in the past
decade, according to the ISI
Essential Science Indicators
Web product.
Dr. McClelland’s
citation record includes 17 papers cited a total of 1,138
times to date in this field. Presently, he is the Walter Van
Dyke Bingham Professor of Psychology and Computer Science at
Carnegie Mellon University, Adjunct Professor of Neuroscience
and the Center for Neuroscience at the University of
Pittsburgh, and the Co-director of the Center for the Neural
Basis of Cognition at Carnegie Mellon. Dr. McNaughton’s
citation record includes 66 papers cited a total of 3,293
times to date in the field of Neuroscience & Behavior. He
holds joint appointments in Psychology and the Neuroscience
Program at the University of Arizona. Dr. O’Reilly’s
citation record includes 10 papers cited a total of 632 times
to date in the field of Psychiatry/Psychology. He is an
Associate Professor in Psychology at the University of
Colorado’s Institute of Cognitive Science, which is part of
their Center for Neuroscience.
|
Why do you think your paper is highly cited?
The paper represents a synthesis of ideas and findings from the
fields of cognitive neuropsychology, computational modeling of
cognition, and systems neuroscience. It provides both a mechanistic
and a functional explanation for the striking patterns of deficits
seen after damage to the hippocampus and related brain areas. It has
been known since the early 1950s that damage to these areas produced
a profound disturbance of the ability to form new memories while
leaving pre-morbid knowledge and cognitive abilities intact. This
striking deficit has been the springboard for a very wide range of
investigations ever since, and so there is considerable interest in
the topic.
What are the circumstances which led you to your work?
One of us (James McClelland) was previously involved in the
development of neural models of other aspects of human cognitive
function. The models McClelland was working with made use of neural
networks that gradually developed cognitive competencies from
extensive experience, capturing the gradual nature of development of
cognitive abilities seen in the early years of life. These models
have been successful in addressing the acquisition of the
"concept" of object permanence, the taxonomic organization
of conceptual knowledge, and the gradual development of reading
skill during the early school years. Some thoughtful colleagues
asked the question whether these models could also address basic
findings in human memory and found that they faced a specific
problem they termed "catastrophic interference." It seemed
that the only way to resolve the apparent conflict between the
successes of the models on the one hand and this failure on the
other was to imagine that the brain contained two distinct and
complementary learning systems, one subserving the gradual discovery
of the underlying structure of experience, and the other subserving
the rapid learning of specific information at typically investigated
in studies of human memory.
These ideas fit together very nicely with a theory that Bruce
McNaughton had been using to guide his neurophysiological
investigations of spatial learning and memory in the rat
hippocampus. This theory, based on early insights of Donald Hebb and
David Marr, is essentially a mechanistic neural network model of
hippocampal function showing how simple synaptic learning mechanisms
proposed by Hebb coupled with information coding strategies explored
by Marr could underlie memory storage in the hippocampus.
Accordingly McClelland arranged to spend a sabbatical with
McNaughton at the University of Arizona learning about the
neuroscience of learning and memory at the cellular and neural
systems levels and working out the details of the theory and
exploring its consistency with the available data from neuroscience.
Conversations with Lynn Nadel at Arizona also had a considerable
influence on the formulation and breadth of coverage of our theory.
After the sabbatical in Arizona, McClelland returned to Pittsburgh
and brought Randy O'Reilly, then a graduate student at Carnegie
Mellon, into the project. Together McClelland and O'Reilly worked on
modeling the inner workings of the hippocampal system based on the
ideas of Hebb, Marr, and McNaughton. This work (published
separately) played a key role by fleshing out some of the details of
the overall theory, and informed the latter stages of developing the
formulation of the overall theory that appears in our three-way
publication.
Would you describe the significance of this work for your
field?
A key element of the theory we offer is that it is a
psychological theory in the sense that it addresses patterns of
deficits seen in tests of human memory, while at the same time the
mechanisms proposed are explicitly neural mechanisms, so that the
theory provides a direct and explicit link between the psychological
and neural levels of analysis. It also helps to link widely
disparate literatures on adult memory, infant and child development,
and the neuropsychology and neurophysiology of learning and memory.
Where has this research gone since the publication of your
paper?
Each of the authors has continued to work on related issues,
pursuing slightly different directions. Around the time we were
developing the theory, McNaughton and his student Matt Wilson
provided the first compelling evidence, predicted by our theory and
the earlier Hebb-Marr theory, that patterns of neural activity
established during a learning episode were re-capitulated
subsequently during sleep. McNaughton and his collaborators have
continued this work, now recording simultaneously from on the order
of 100 neurons at a time in the brains of rats and monkeys, allowing
them to track the patterns of neural activity laid down during
experience and reactivated at a later time. In particular, they are
now focusing on the statistical interactions between the hippocampus
and neocortex during sleep-related memory trace reactivation that
are predicted by the theory.
O'Reilly has built extensively on this theory, applying it to a
wide range of findings in both the human and animal learning and
memory literatures. Among other work, O'Reilly has collaborated with
others to produce two subsequent articles in Psychological Review
that build directly on the theory laid out in our 1995 paper.
McClelland's subsequent work has focused on the further
development of the theory of how gradual learning in the neocortex
gives rise to the emergence of conceptual knowledge in childhood and
subserves advanced adult cognitive abilities in tasks that depend on
conceptual knowledge. Together with a former student, McClelland now
has a book in press on this topic (Rogers, T. T. and McClelland, J.
L. Semantic Cognition: A Parallel Distributed Processing
Approach. Cambridge, MA, MIT Press, to appear in 2004).
Where do you see it going 10 years from now?
In spite of the fact that this work has struck a responsive
chord, there is a great deal we do not yet understand about the
nature of learning and memory. We can dream of a day—how soon it
will come we don't know—when we will have a fully elaborated
theory of the mechanisms of learning and memory in the human (and
non-human mammal) brain. While the three of us will continue to
contribute to this effort, it is one that is also being pursued by
many others as well, including individuals working at the molecular,
cellular, systems, computational, and behavioral levels. It is
through the integration of research at all these levels that we will
continue to make progress toward the goal of understanding the
mechanistic basis of learning and memory.
What lessons would you draw from your
work to share with the next generation of researchers?
The success of our work owes itself, we think, to two things: (1)
Taking problems raised by others (in this case, the catastrophic
interference problem) seriously, and (2) seeking to build bridges
between levels of analysis (in this case, the behavioral level and
the neural level).
James L. McClelland, Ph.D.
Carnegie Mellon University
Pittsburgh, PA, USA
Bruce McNaughton, Ph.D.
University of Arizona
Tucson, AZ, USA
Randall C. O’Reilly, Ph.D.
University of Colorado
Boulder, CO, USA
|
cites
