2005 Special issue
Methods for reducing interference in the Complementary Learning Systems
model: Oscillating inhibition and autonomous memory rehearsal
Kenneth A. Norman*, Ehren L. Newman, Adler J. Perotte
Department of Psychology Princeton University, Green Hall, Princeton, NJ 08544, USA
The stability–plasticity problem (i.e. how the brain incorporates new information into its model of the world, while at the same time preserving
existing knowledge) has been at the forefront of computational memory research for several decades. In this paper, we critically evaluate how well
the Complementary Learning Systems theory of hippocampo–cortical interactions addresses the stability–plasticity problem. We identify two
major challenges for the model: Finding a learning algorithm for cortex and hippocampus that enacts selective strengthening of weak memories,
and selective punishment of competing memories; and preventing catastrophic forgetting in the case of non-stationary environments (i.e. when
items are temporarily removed from the training set). We then discuss potential solutions to these problems: First, we describe a recently
developed learning algorithm that leverages neural oscillations to find weak parts of memories (so they can be strengthened) and strong
competitors (so they can be punished), and we show how this algorithm outperforms other learning algorithms (CPCA Hebbian learning and
Leabra at memorizing overlapping patterns. Second, we describe how autonomous re-activation of memories (separately in cortex and
hippocampus) during REM sleep, coupled with the oscillating learning algorithm, can reduce the rate of forgetting of input patterns that are no
longer present in the environment.We then present a simple demonstration of how this process can prevent catastrophic interference in an AB–AC
q 2005 Elsevier Ltd. All rights reserved.
Keywords: Hippocampus; Neocortex; Neural network; Interference; Learning algorithm; Theta Oscillations; Sleep; Consolidation
Over the past several decades, neural theorists have
converged on the idea that neocortex implements an internal,
predictive model of the structure of the environment. This
internal model must simultaneously maintain previously
learned information and integrate new information. The
problem of how to accomplish these goals simultaneously in
a neural network architecture was labeled the stability–
plasticity dilemma by Carpenter and Grossberg (1988), and
this problem has come to occupy a central position in
computational neuroscience. The problem is hard to solve
because, in most neural network models, memory traces
overlap with one another. As such, learning new memories will
incrementally degrade pre-existing memories. Several
researchers have found that, when new learning is extensive
(e.g. if the system has to memorize a new pattern based on a
single learning trial), neural networks can show near-complete
forgetting of pre-existing knowledge (catastrophic interfer-
ence; French, 1999; French, 2003; McCloskey & Cohen,
There have been several attempts to solve this problem, e.g.
Adaptive Resonance Theory (Carpenter & Grossberg, 2003]).
In this paper, we focus on another framework for addressing
stability–plasticity: The Complementary Learning Systems
(CLS) model (McClelland, McNaughton, & O’Reilly, 1995;
O’Reilly & Norman, 2002; O’Reilly & Rudy, 2001; Norman &
O’Reilly, 2003). This model posits that cortex solves stability–
plasticity with the assistance of a hippocampal system that can
rapidly memorize events and play them back to cortex in an
‘off-line’ fashion. In Section 1.2, we describe the basic
properties of CLS, and how it is meant to solve stability–
We also briefly review some of the many ways in which
CLS has been applied to episodic memory and animal learning
data. However, while CLS has proved to be a very useful way
for thinking about hippocampal and cortical learning
processes, in recent years we have identified some issues
with the model that we want to address:
Neural Networks 18 (2005) 1212–1228
0893-6080/$ - see front matter q 2005 Elsevier Ltd. All rights reserved.
E-mail addresses: email@example.com (K.A. Norman), enewman@
princeton.edu (E.L. Newman), firstname.lastname@example.org (A.J. Perotte).
† The first issue involves finding a suitable algorithm for
adjusting synapses in cortex and the hippocampus. Some of
the learning algorithms that have been used in CLS
implementations (e.g. CPCA Hebbian learning: Norman
& O’Reilly, 2003; O’Reilly & Munakata, 2000) adjust
synapses more than is necessary and, as such, show
unacceptably high levels of interference. Other learning
rules that have been used in CLS implementations (e.g.
Leabra; O’Reilly & Munakata, 2000) are less prone to this
problem, but have other problems of their own (e.g. both
Leabra and CPCA Hebbian learning have difficulty in
modeling data on how competitors are punished during
† The second issue involves the problem of non-stationary
environments: What happens when patterns that were
originally in the training set are removed from the
training set? Even with the hippocampus and cortex
working together, the standard form of the CLS model
shows unacceptably high rates of forgetting of patterns
once they are removed from the training set. This
problem needs to be addressed before the CLS model
can be viewed as a complete solution to the stability–
In this paper, we present solutions to both of these problems:
† In section 2, we describe a new learning algorithm
developed by Norman, Newman, Detre, and Polyn (2005)
that leverages regular oscillations in feedback inhibition to
pinpoint weak parts of target memories (so they can be
strengthened) and to pinpoint non-target memories that
compete with target memories during retrieval (so they can
be weakened). We show that the oscillating learning
algorithm, applied to our cortical network, outperforms
both CPCA Hebbian learning and Leabra on a pattern
completion task. We also show that the oscillating learning
algorithm’s capacity for supporting familiarity discrimi-
nation greatly exceeds the capacity of the Hebbian cortical
model from Norman and O’Reilly (2003).
† In section 3, we show how the CLS model can be
supplemented by a new kind of off-line learning where
cortex and hippocampus separately rehearse stored mem-
ories, thereby repairing damage to these memories. We
argue that this off-line learning reflects the functionality of
REM sleep, and show that it can successfullyprevent loss of
knowledge in an AB–AC interference paradigm (where AB
items are initially trained and then removed from the
In summary: We will present an account of how inhibitory
oscillations and off-line rehearsal of stored knowledge (during
REM sleep) can both improve learning and retention. The ideas
presented here apply to both hippocampus and cortex. For
simplicity’s sake, the simulations that we present will use the
cortical model, which has a less differentiated architecture than
the hippocampal model. After each simulation, we will discuss
ways in which the same mechanism can be applied to the
1.2. Basic properties of CLS
The CLS framework (McClelland et al., 1995) incorporates
several widely-held ideas about hippocampal and neocortical
contributions to memory, that have been developed over many
years by many different researchers (e.g. Aggleton & Brown,
1999; Burgess & O’Keefe, 1996; Eichenbaum, Otto, & Cohen,
1994; Grossberg, 1976; Hasselmo & Wyble, 1997; Marr, 1971;
McNaughton & Morris, 1987; Moll & Miikkulainen, 1997;
O’Keefe & Nadel, 1978; Rolls, 1989; Scoville & Milner, 1957;
Sherry & Schacter, 1987; Squire, 1992; Sutherland & Rudy,
1989; Teyler & Discenna, 1986; Treves & Rolls, 1994; Wu,
Baxter, & Levy, 1996; Yonelinas, 2002). According to the CLS
framework, neocortex forms the substrate of our internal model
of the structure of the environment. In contrast, hippocampus is
specialized for rapidly and automatically memorizing patterns
of cortical activity, so they can be recalled later (based on
The CLS framework posits that neocortex learns incremen-
tally; each training trial results in relatively small adaptive
changes in synaptic weights. These small changes allow cortex
to gradually adjust its internal model of the environment in
response to new information. The other key property of
neocortical learning is that it assigns similar (overlapping)
representations to similar stimuli. Use of overlapping
representations allows cortex to represent the shared structure
of events, and therefore makes it possible for cortex to
generalize to novel stimuli based on their similarity to
previously experienced stimuli. In contrast, hippocampus is
biased to assign distinct, pattern separated representations to
stimuli, regardless of their similarity. This property allows
hippocampus to rapidly memorize arbitrary patterns of cortical
activity without suffering catastrophic levels of interference.
1.3. How CLS solves stability–plasticity
One of the key problems facing any account of stability–
plasticity is how to incorporate rare (but significant) events into
the cortical network. In the case of the CLS model, the
incremental nature of cortical learning means that it can only
retrieve memories if the stimulus is presented repeatedly.
However, infrequently-occurring events are sometimes very
significant (e.g. if a pterodactyl eats your sister) and we need to
be able to incorporate this information into our internal cortical
model of how the world works, so we can properly generalize
to new situations (e.g. future pterodactyl attacks). If the cortical
network were left to its own devices, a person would have to
experience several pterodactyl attacks before the cortical
memory trace was strong enough to support appropriate recall
and generalization. Furthermore, if the average interval
between pterodactyl appearances were sufficiently long, one
runs the risk that—in between appearances—interference from
other memories would erode the original memory, in which
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281213
case the person would be back to where they started with each
new pterodactyl appearance.
The presence of the hippocampal network solves this
problem. The hippocampus is specialized for rapid memoriza-
tion; in a single trial, the hippocampus can latch on to pattern of
cortical activity elicited by the pterodactyl, and re-play it to
cortex repeatedly until it sinks in. In this respect, hippocampus
can be viewed as a ‘training trial multiplier’. Over time,
hippocampally-mediated replay of pterodactyl memories is
interleaved with bottom-up learning about information in the
environment. As discussed by McClelland et al. (1995), this
kind of interleaved training, coupled with a learning mechan-
ism that is sensitive to prediction error, forces cortex to develop
representations that reconcile the properties of rare events and
more common events (because this is the only way to avoid
prediction error across the entire training set).1
1.4. Applications of CLS to episodic memory and other
CLS was originally formulated as a set of high-level
principles for understanding hippocampal and cortical contri-
butions to memory. More recently, O’Reilly and Rudy (2001)
and Norman and O’Reilly (2003) have developed working
neural network models of hippocampus and neocortex that
instantiate these principles, and these networks have been
applied to modeling specific datasets.
1.5. Modeling recognition memory
In one application, Norman and O’Reilly (2003)
implemented hippocampal and cortical networks that adhere
to CLS principles, and showed how these networks (taken
together) constitute a neural dual-process model of recognition
memory. Learning was implemented in these simulations using
a simple Hebbian rule (called instar learning by Grossberg,
1976, and CPCA Hebbian learning by O’Reilly & Munakata,
2000), whereby connections between active sending and
receiving neurons are strengthened, and connections between
active receiving neurons and inactive sending neurons are
weakened. Norman and O’Reilly (2003) showed how the
hippocampal model (using this simple Hebbian rule) can
support recognition via recollection of specific studied details.
The cortical model cannot support recollection of specific
details from once-presented events, owing to its relatively low
learning rate. However, Norman and O’Reilly (2003) showed
that cortex can still support judgments of familiarity after a
single study trial, based on the sharpness of representations in
The cortical model’s ability to support familiarity discrimi-
nation is a simple consequence of Hebbian learning and
inhibitory competition. When a stimulus is presented, Hebbian
learning tunes a subset of the hidden units to respond more
strongly to that stimulus. As these units respond more and more
strongly to the stimulus, they start to inhibit other units. Thus,
the neural response to a stimulus transitions from a diffuse
overall response (where no units are tuned to respond strongly
to the stimulus) to a more focused response where some units
are strongly active and other units are suppressed. In the
Norman and O’Reilly (2003) paper, cortical familiarity was
operationalized in terms of the activation of the k most active
units in the hidden layer (where k is a model parameter that
defines the maximum number of units that are allowed to be
strongly active at once), although other methods of operatio-
nalizing familiarity are possible.
Norman and O’Reilly (2003) showed how, taken together,
the hippocampal network and cortical network could explain a
wide range of recognition findings, including data on when
hippocampal lesions affect recognition memory (as a function
of how similar distractors are to studied items, and as a function
of test format) and data from normal subjects on how
interference manipulations affect recognition memory (e.g.
list strength manipulations: how does repeatedly presenting
some items on the study list affect memory for other items on
the study list).
1.6. Modeling animal learning
In another application, O’Reilly and Rudy (2001) used
hippocampal and cortical networks instantiating CLS prin-
ciples to explain findings from animal learning paradigms,
including non-linear discrimination learning (e.g. negative
patterning, transverse patterning), ‘transitive inference’ in
discrimination learning, and contextual fear conditioning.
The models in these simulations were largely identical to the
models used in Norman and O’Reilly (2003), except the
simulations used O’Reilly’s Leabra learning rule instead of
CPCA Hebbian learning. Leabra combines CPCA Hebbian
learning with a simple form of error-driven learning (O’Reilly
& Munakata, 2000). The key finding from these simulations
was that cortex could solve non-linear discrimination problems
on its own when the animal is given repeated exposure to the
stimuli and appropriate feedback. In contrast, hippocampus is
needed to show sensitivity to feature conjunctions on tasks
where conjunctive learning is incidental (i.e. the animal does
not have to learn the conjunction to respond correctly on the
task) and the animal is given limited exposure to the
conjunction. O’Reilly and Rudy (2001) discuss several findings
that support the model’s predictions.
1.7. Problems with learning rules
Concrete applications of CLS (like those described in
Norman & O’Reilly, 2003 and O’Reilly & Rudy, 2001) have
1One could reasonably ask why we need to represent rare events in cortex,
given that hippocampus is capable of recalling these events after a single
training trial. The answer (according to CLS) is that, even though hippocampus
can support recall, it is not well suited to feature-based generalization. Thus, to
the extent that we want to generalize properly to similar events in the future
(e.g. different colors and sizes of pterodactyls appearing in different locations),
information about pterodactyls needs to be transferred from hippocampus to
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281214
provided strong support for the validity of basic CLS principles
(see also O’Reilly & Norman, 2002). However, the process of
building working models that instantiate CLS principles has
also highlighted some important challenges for the CLS
One critical challenge is to develop a learning algorithm that
is capable of storing an appropriately large database of
knowledge (semantic knowledge, in the case of cortex, and
episodic knowledge, in the case of the hippocampus). Norman
and O’Reilly (2003) noted that the CPCA Hebbian learning
rule used in that paper has a tendency to over-focus on
prototypical features. When given a large set of correlated
input patterns to memorize, the CPCA Hebbian algorithm is
very good at learning what all of these patterns have in
common, but it shows very poor memory for specific, non-
prototypical features of individual items. This is less of a
problem for the hippocampal model than forthe corticalmodel,
because of the hippocampal model’s ability to assign relatively
distinct representations to similar inputs. However, Norman
and O’Reilly (2003) noted that the hippocampal model is still
prone to ‘pattern separation collapse’ when given large
numbers of overlapping patterns. When this occurs, the
hippocampus recalls prototypical features in response to all
input patterns (studied or non-studied).
From a psychological-modeling perspective, the mere fact
that Hebbian learning over-focuses on prototypes is not
problematic. Good memory for prototypes can be used to
explain numerous categorization and memory phenomena (e.g.
false recognition of non-studied items from studied categories;
Koutstaal, Schacter, & Jackson, 1999). Also, as discussed by
Norman and O’Reilly (2003), the model’s tendency to forget
individuating features of studied items can be used to explain
memory interference effects on list learning paradigms.
However, the excessive degree of prototype-focusing
exhibited by the model is more problematic. When the model
is given a sufficiently large number of overlapping patterns,
both the hippocampal and cortical networks exhibit virtually no
memory for individuating features. In an important analysis,
Bogacz and Brown (2003) set out to quantify the capacity of
several different cortical models (including the Norman &
O’Reilly, 2003 Hebbian cortical network) for supporting
familiarity-based recognition: How many patterns can be
stored in the network, in a manner that supports discrimination
of studied vs. non-studied patterns? This analysis showed that,
given overlapping input patterns, the capacity of the Hebbian
cortical network from Norman and O’Reilly (2003) was very
poor. Even in a brain-sized version of the network, the model’s
capacity is almost certainly not large enough to account for
data on human recognition memory capacity (e.g. Standing,
1973) showed that people can discriminate between thousands
of studied vs. non-studied pictures, and this is an extremely
1.8. Why does CPCA Hebbian learning perform poorly?
The essence of the problem with CPCA Hebbian learning is
that it is insufficiently judicious in how it adjusts synaptic
strengths. In neural networks, each synaptic weight is involved
in storing multiple memories. As such, adjusting weights to
improve recall of one memory interferes with other memories
that are encoded in those weights. Given that there is a cost (in
terms of interference) as well as a benefit to adjusting synaptic
weights, it makes sense that strengthening of weights should
stop once the target memory is strong enough to support recall
and generalization. Likewise, learning algorithms should only
weaken non-target memories that are actively competing with
recall of the target memory. Any further strengthening (of the
target memory) or weakening (of non-target memories) will
cause interference without improving recall. CPCA Hebbian
learning fails on both counts: It strengthens synapses between
co-active units even if the target memory is already strong
enough to support recall, and it weakens synapses between
active receiving units and all sending units that are inactive at
the end of the trial, even if these units did not actively compete
with recall of the target memory.
In addition to being inefficient (from a functional
standpoint), CPCA Hebbian learning’s inability to selectively
weaken competing memories also impedes its ability to
account for empirical data on competitor punishment. Over
the past decade, several studies have found that memory
weakening is modulated by how strongly memories compete at
retrieval: Non-target memories that compete strongly with the
target memory (but subsequently lose the competition to be
retrieved) are punished. However, if steps are taken to mitigate
competition (e.g. by increasing the specificity of the retrieval
cue), there is no punishment (see Anderson, 2003 for a review
of these findings; see also Norman, Newman, & Detre, 2004 for
a computational model of these findings). This pattern of
results has been observed in both semantic memory tasks (e.g.
Blaxton & Neely, 1983) and episodic memory tasks (e.g.
Anderson & Bell, 2001; Ciranni & Shimamura, 1999),
suggesting that selective competitor punishment occurs in
both cortex and hippocampus. However, contrary to these
findings, CPCA Hebbian learning predicts that all memories
that overlap with the target memory should be weakened,
regardless of the amount of competition at retrieval.
1.9. Problems with Leabra
As mentioned earlier, some implementations of CLS (e.g.
O’Reilly & Rudy, 2001) have used O’Reilly’s Leabra learning
algorithm instead of CPCA Hebbian learning. Because of its
ability to learn based on pattern completion error, Leabra does
a much better job than CPCA Hebbian learning at retaining the
individuating features of studied items. However, as discussed
in Norman et al. (2004), Leabra lacks a mechanism for
selectively punishing memories that compete at retrieval. The
essence of this problem is that competitor activity is transient
(i.e. the competitor ‘pops up’ briefly and then goes away), but
Leabra is only equipped for learning about representations that
are active in the final settled state of the network. As such,
Leabra also fails to account for the competitor-punishment data
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281215
1.10. Desiderata for a replacement algorithm
Because of the issues with CPCA Hebbian learning and
Leabra outlined above, we set out to derive a new learning
algorithm that meets the following two desiderata:
† Limits on strengthening: The network should only
strengthen memories when they are too weak to support
† Targeted punishment: The network should only weaken
memories when they actively compete with successful
recall of the target memory.
These properties, taken together, should reduce interference
in the cortical and hippocampal models. The second property
should help the networks account for data on competitor
2. The oscillating learning algorithm
To meet the desiderata outlined above, Norman et al. (2005)
developed a new learning algorithm that selectively strength-
ens weak parts of target memories (vs. parts that are already
strong), and selectively punishes strong competitors. The
learning algorithm accomplishes this goal by oscillating the
strength of feedback inhibition, and learning based on
the resulting changes in activation. In this section, we first
provide some background information on how inhibition was
implemented in the model, and how the network was
structured. We then provide a highlevel overview of how the
algorithm works. Finally, we present benchmark data (taken in
part from Norman et al., 2005) comparing the oscillating
learning algorithm to Leabra and CPCA Hebbian learning.
2.1. Background: how inhibition was implemented in the model
In the simulations described below, we used the simple two-
layer cortical network shown in Fig. 1. The network was
provided with patterns to memorize on the input/output layer,
and the hidden layer was free to self-organize. Every
input/output unit was connected to every input/output unit
(including itself) and to every hidden unit. Whenever a network
is recurrently connected in this manner, there has to be some
mechanism for limiting the spread of excitatory activity. In the
brain, this problem is solved by inhibitory interneurons, which
enforce a set point on the amount of excitatory activity within a
subregion (O’Reilly & Munakata, 2000). We capture this set
point dynamic in our model using a k-winners-take-all (kWTA)
inhibition rule, which adjusts inhibition such that the k units in
each layer that receive the most excitatory input are strongly
active, and all other units are at most weakly active (activity !
.25; Minai & Levy, 1994; O’Reilly & Munakata, 2000). We set
the input/output layer k equal to the number of units in each
studied pattern, such that (when kWTA is applied to the
network) the best-fitting memory—and only that memory—is
2.2. Algorithm overview
The learning algorithm can be described in the following
† First,the target pattern is soft-clamped onto the input/output
layer of the network. This soft-clamp is applied for the
duration of the trial. Given normal levels of inhibition, the
kWTA rule prevents activation from spreading to other
units in the input/output layer.
† Second, the algorithm identifies competitors by lowering
inhibition below the level specified by kWTA. Effectively,
lowering inhibition lowers the threshold amount of
excitation needed for a unit to become active. If a non-
target unit is just below threshold (i.e. it is receiving strong
input, but not quite enough to become active) lowering
inhibition will cause that unit to become active.
† Third, the algorithm weakens units that turn on when
inhibition is lowered (i.e. strong competitors) by reducing
weights from other active units. By doing this, the learning
algorithm ensures that a unit that competes on one trial will
receive less input the next time that cue is presented. If the
unit will diminish to the point where it no longer activates in
the low inhibition condition (so no further punishment
occurs). Norman et al. (2004) describe how this property
allows the model to simulate detailed patterns of behavioral
data on competitor-punishment.
† Fourth, the algorithm identifies weak parts of target
memories by raising inhibition above the level specified
by kWTA. If a target unit is receiving relatively little
Fig. 1. Diagram of the network used in our simulations. Patterns were presented
to the lower part of the network (the input/output layer). The upper part of the
network (the hidden layer) was allowed to self-organize. Every unit in the
input/output layer was connected to every input/output unit (including itself)
and to every hidden unit via modifiable, symmetric weights.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281216
collateral support from other target units, such that its net
input is just above threshold, raising inhibition will trigger a
decrease in the activation of that unit.
† Fifth, the algorithm strengthens units that turn off when
inhibition is raised (i.e. weak target units) by increasing
weights from other active units. By doing this, the learning
algorithm ensures that a target unit that drops out on a given
trial will receive more input the next time that cue is
presented. If the same pattern is presented repeatedly,
eventually the input to that unit will increase to the point
where it no longer drops out in the high inhibition condition
(so no further strengthening occurs).
2.3. Algorithm details
The algorithm uses Contrastive Hebbian Learning (CHL;
Ackley, Hinton, & Sejnowski, 1985; Hinton, 1989; Hinton &
McClelland, 1988; Hinton & Sejnowski, 1986; Movellan,
1990) to enact the weight changes described above. CHL
involves contrasting a more desirable state of network activity
(the plus state) with a less desirable state of network activity
(the minus state). The CHL equation adjusts network weights
so that the more desirable state of network activity is more
likely to occur in the future. The following equation shows how
weight changes are computed by CHL:
In the above equation, Xiis the activation of the presynaptic
(sending) unit, Yj is the activation of the postsynaptic
(receiving) unit. The ‘C’ and ‘K’ superscripts refer to plus-
state and minus-state activity, respectively. dWijis the change
in weight between the sending and receiving units, and 3 is the
learning rate parameter.
The oscillating learning algorithm generates minus states by
varying inhibition around the level set by kWTA. When
inhibition is at its normal level (i.e. the level set by kWTA), all
of the target units (and only those units) will be active. This is
the maximally correct state of network activity. On each trial,
we distort this pattern by oscillating inhibition in a continuous
fashion from its normal level to lower-than-normal, then to
higher-than-normal, then back to normal, and we apply the
CHL equation to successive time steps of network activity. At
each point in the inhibitory oscillation, inhibition is either
moving toward its normal level (the ‘maximally correct’ state),
or it is moving away from this state. If inhibition is moving
toward its normal level, then the activity pattern at time tC1
will be more correct than the activity pattern at time t. In this
situation, we use the CHL equation to adapt weights to make
the pattern of activity at time t more like the pattern at time tC
1. However, if inhibition is moving away from its normal level,
then the activity pattern at time tC1 will be less correct than
the activity pattern at time t (it will either contain too much or
too little activity, relative to the target pattern). In this situation,
we use the CHL equation to adapt weights to make the pattern
of activity at time tC1 more like the pattern at time t. These
rules are formalized in Eqs. (2) and (3).
If inhibition is returning to its normal value:
If inhibition is moving away from its normal value:
For a detailed description of how the algorithm was
implemented, see Norman et al. (2005).
2.4. Relation to neural oscillations
Although the algorithm was not specifically developed as a
theory of neural oscillations, it nonetheless may help to explain
how neural oscillations contribute to learning. In particular,
theta oscillations (rhythmic changes in local field potential at a
frequency of approximately 4–8 Hz in humans) have several
properties that resonate with the learning algorithm proposed
† Theta oscillations depend critically on changes in the firing
of inhibitory interneurons (Buzsaki, 2002; Toth, Freund, &
† Theta oscillations have been observed in both of the major
CLS structures (cortex and hippocampus; for a review, see
Kahana, Seelig, & Madsen, 2001).
† Theta oscillations are fast enough to support several
complete oscillations per stimulus presentation, and slow
enough to allow competitors to activate when inhibition is
† Theta oscillations have been linked to learning, in both
animal and human studies (e.g. Raghavachari et al., 2001).
Several studies have found that the direction of potentiation
(LTP vs. LTD) depends on the phase of theta (peak vs.
trough; Holscher, Anwyl, & Rowan, 1997; Huerta &
Lisman, 1996; Hyman, Wyble, Goyal, Rossi, & Hasselmo,
2003). This result mirrors the property of our model
whereby the high-inhibition phase of the oscillation is
primarily concerned with strengthening target memories
(LTP) and the low-inhibition phase of the oscillation is
primarily concerned with weakening competitors (LTD).
Given these facts, it seems possible to us that theta
oscillations may serve as the neural substrate of the algorithm
described here (Norman et al., 2005). However, at this point the
linkage is only suggestive, and needs to be confirmed through
2.5. Pattern completion simulations
To explore the oscillating algorithm’s ability to avoid
pattern separation failure and recall individuating features,
Norman et al. (2005) ran simulations comparing pattern
completion performance for the oscillating algorithm vs.
Leabra. In one set of simulations, Norman et al. (2005) gave
the network 200 binary input patterns to learn, with 57%
K.A. Norman et al. / Neural Networks 18 (2005) 1212–1228 1217
average overlap between patterns. The network was repeatedly
presented with the 200-pattern set until learning reached
asymptote. At the end of training, pattern completion was
tested by measuring the network’s ability to recall a single,
non-prototypical feature from each pattern, given all of the
other features of that pattern as a retrieval cue.
Norman et al. (2005) were also interested in comparing the
robustness of the representations learned by each algorithm:
To what extent can these representations support retrieval
when test cues do not exactly match studied patterns? To get at
this issue, they distorted retrieval cues (by adding Gaussian
noise to the input pattern that was clamped onto the network)
and measured how pattern completion performance varied as a
function of the amount of test-pattern noise.
Fig. 2 shows the number of patterns (out of 200)
successfully recalled at the end of training by the oscillating
learning algorithm and Leabra, as a function of the amount of
noise applied to retrieval cues at test (Norman et al., 2005).
For comparison purposes, we have also included the results of
new simulations using CPCA Hebbian learning. In keeping
with the idea (stated earlier) that CPCA Hebbian has difficulty
in learning about non-prototypical features, this algorithm
performed very poorly, even for low levels of noise. Leabra
performed better than CPCA Hebb; this is because the error-
driven component of Leabra explicitly computes pattern
completion error at training, and adjusts weights to reduce
this error. When test noise was set to zero, Leabra and the
oscillating algorithm performed comparably. However, when
the models were given noisy test cues, the oscillating algorithm
performed much better than Leabra.
The oscillating learning algorithm outperforms Leabra in
this situation because the oscillating algorithm does a better job
of maintaining pattern separation in the hidden layer: At the
end of training, the average pairwise similarity between
patterns in the hidden layer (measured using cosine) was .47
(SEMZ.02) for the oscillating algorithm vs. .65 (SEMZ.01)
for Leabra. The high level of hidden-layer overlap in the
Leabra simulations hurts recall by increasing the odds that
(given a noisy input pattern) the network will slip out of the
correct attractor into a neighboring attractor. The oscillating
learning algorithm manages to avoid these pattern-separation
difficulties because of its ability to punish competitors:
Whenever memories start to blend together, they also start to
compete with one another at retrieval, and the competitor-
punishment mechanism pushes them apart.
Crucially, even though pattern separation is higher for the
oscillating learning algorithm vs. Leabra, the oscillating
algorithm still learns similarity-based representations (i.e. it
assigns similar hidden representations to similar inputs). The
quantify this tendency, we computed a ‘similarity score’ that
tracks the correlation (across all pairs of patterns) between
input-layer similarity and hidden-layer similarity. The average
similarity score for the oscillating algorithm was .58 (SEMZ
.02), vs. .71 (SEMZ.04) for Leabra. Although the mean
similarity score for Leabra was higher, similarity scores for
Leabra were also much more variable: Across runs, some
scores were extremely high, and some scores were extremely
low. Approximately 10% of the Leabra similarity scores were
less than .1, indicating a near-total failure to represent the
structure of the input space. In contrast, only two runs of the
oscillating-algorithm model yielded similarity scores below .5,
and these scores (.35 and .45) still showed substantial
sensitivity to the structure of the input space.
Finally, the fact that oscillating algorithm learns similarity-
based representations (given a cortical network architecture)
highlights an essential difference between the pattern separ-
ation mechanisms that are built into the CLS hippocampal
model, and the pattern separation enacted by the oscillating
algorithm. As discussed earlier, the goal of hippocampal
pattern separation is to assign maximally distinct represen-
tations to stimuli, regardless of their similarity, so these stimuli
can be recalled in detail. The hippocampal model’s extreme
approach to pattern separation effectively cripples its ability to
generalize. In contrast, the oscillating algorithm is only
concerned that memories observe a ‘minimum separation’
from one another. So long as this constraint is met, memories in
the cortical network simulated here are free to overlap
according to their similarity (thereby allowing the network to
enact similarity-based generalization).
2.6. Familiarity discrimination: comparison with Hebbian
In addition to the pattern completion simulations described
in Norman et al. (2005), we have recently started to use the
oscillating algorithm to simulate familiarity-based recognition
in the cortical network. It is possible to read out a familiarity
score from the oscillating algorithm by looking at how
activation changes when inhibition is raised above its baseline
Patterns Learned as a Function of Test Pattern Noise
Test Pattern Noise (x 10-2)
02468 10 12 14 16
Number of Patterns Learned
Leabra (Hebb .01)
Fig. 2. Comparison of pattern completion performance for the oscillating
learning algorithm vs. other learning algorithms. The figure shows the number
of patterns (out of 200) successfully recalled at the end of training by each
algorithm, as a function of the amount of noise applied to retrieval cues at test;
the oscillating-algorithm and Leabra results are taken from Norman et al.
(2005). CPCA Hebbian learning performs very poorly. The oscillating learning
algorithm and Leabra perform comparably for low noise values, but the
oscillating algorithm performs much better than Leabra for noisy retrieval cues.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281218
value: Weak (unfamiliar) memories show a larger decrease in
activation than strong (familiar) memories.
In new simulation work (not published elsewhere), we
have found that the capacity of the oscillating algorithm for
supporting familiarity discrimination is much higher than the
capacity of the Hebbian familiarity discrimination model used
by Norman and O’Reilly (2003). For example, in one
simulation we generated 200 patterns with 41% average
overlap. We trained the network by presenting 100 of the
patterns for 10 epochs. After each epoch of training, we tested
the network’s ability to discriminate between the 100 patterns
it studied, and the 100 patterns that it did not study. For the
oscillating-algorithm familiarity simulations, we used the
same network that we used in the pattern completion
simulations above (with 80 input units and 40 hidden units).
We compared the results of the oscillating-algorithm
simulations to the results of simulations using the feedforward
Hebbian model from Norman and O’Reilly (2003). The only
change to the Hebbian model as described in that paper is that
we used 80 input units instead of 240. The exact same input
patterns were presented to the oscillating-algorithm model and
the Hebbian model. The Hebbian model simulations oper-
ationalized familiarity using the activation of winners
Familiarity measure that was introduced by Norman and
O’Reilly (2003): familiarity is the average activation of the k
most active hidden units (where k is the activation limit
imposed by the k-winners-take-all inhibition rule). For the
oscillating-algorithm simulations, we used two different
familiarity measures. In one set of simulations, we indexed
familiarity in terms of the change in average activation (over
the entire input layer) given high vs. normal inhibition. Also,
to maximize comparability with the Hebbian simulations, we
ran another set of oscillating-algorithm simulations using the
activation of winners familiarity measure from Norman and
The results of these simulations are shown in Fig. 3. The
Hebbian model performs just above chance after one epoch of
training, and actually gets slightly worse with additional
training. This result is robust to a wide range of parameter
settings, including hidden layer size—it appears that there is
simply no way to get the Hebbian model to show good
discrimination of patterns with this level of overlap. This is
because of the Hebbian model’s tendency to over-focus on
prototype features. In contrast, after 10 epochs, the oscillating
learning algorithm showedO92% accuracy in discriminating
between studied and non-studied patterns using the same
familiarity measure (activation of winners) that was used in the
Hebbian model. Asymptotic accuracy was even better (O99%)
when we used a familiarity measure (change in activation,
given high vs. normal inhibition) that was specifically tailored
to the oscillating algorithm. Although we have not yet carried
out the requisite mathematical analyses, we think it is quite
possible that the oscillating algorithm’s capacity for supporting
familiarity-based discrimination, in a brain-sized network, will
be large enough to account for the vast capacity of human
familiarity discrimination (as illustrated, e.g. by Standing,
2.7. Extending the oscillating algorithm to the hippocampal
The basic principles of the oscillating algorithm (regarding
how changes in the strength of inhibition can be used to
identify weak parts of target memories, and to flush out
competitors) should apply to the hippocampus just as well as
they apply to cortex. However, as discussed by Norman et al.
(2005), our ideas regarding the functional role of theta
oscillations differ from other published theories of how theta
contributes to hippocampal processing. Most prominently,
Hasselmo, Bodelon, and Wyble (2002) have argued that theta
oscillations help tune hippocampal dynamics for encoding vs.
retrieval, such that dynamics are optimized for encoding during
one phase of theta, and dynamics are optimized for retrieval
during another phase of theta. The Hasselmo et al. (2002)
model varies the relative strengths of different excitatory
projections as a function of theta (to foster encoding vs.
retrieval), but does not vary inhibition. Our impression is that
our oscillating algorithm and Hasselmo’s model are orthogonal
rather than contradictory. As such, we may be able to combine
the twomodels.One possibility would be to align the inhibitory
oscillation (from our model) and the oscillation in excitatory
projection strengths (from Hasselmo’s model) such that
inhibition is above-baseline during the ‘encoding’ phase of
theta and inhibition is below-baseline during the ‘retrieval’
phase of theta. As per our theory, learning would be based on
changes in activation triggered by changing inhibition. This
Familiarity Discrimination in the Old and New Models
Epochs of Training
Oscillating algorithm: Activation of winners
Oscillating algorithm: Change in activation
Hebbian model: Activation of winners
Fig. 3. Comparison of familiarity discrimination using the oscillating learning
algorithm vs. the Norman and O’Reilly (2003) Hebbian cortical familiarity
model. For the Hebbian model, familiarity was operationalized using the
activation of winners measure from Norman and O’ Reilly (2003). For the
oscillating-algorithm model, familiarity was operationalized in two different
ways: activation of winners, and also the change in activation given high vs.
normal inhibition. Note that the oscillating-algorithm simulations used 40
hidden units, whereas the Hebbian simulations used 1920 hidden units (to
match the simulations from Norman & O’Reilly, 2003). Despite this large
disparity in hidden layer size, the oscillating-algorithm familiarity model
strongly outperformed the Hebbian model: Given 100 patterns (and 41%
average overlap between patterns), the asymptotic accuracy of the oscillating-
algorithm simulations was O99% for the change in activation measure and O
92% for the activation of winners measure, whereas the Hebbian model’s
asymptotic accuracy was close to chance.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281219
method of lining up the oscillations has the useful property that
the oscillation phase primarily associated with memory
strengthening in our model (high inhibition) matches up with
the oscillation phase associated with LTP in the Hasselmo
model (‘encoding mode’), and the oscillation phase primarily
associated with competitor punishment in our model (low
inhibition) matches up with the oscillation phase associated
with LTD in the Hasselmo model (‘retrieval mode’). We will
explore the viability of this combined model in future research.
3. Model of memory protection during REM
The preceding section focused on problems with the
learning rules used by CLS models, and how these problems
might be addressed using the oscillating learning algorithm.
However, there are other, deeper issues with the CLS
framework that cannot be addressed simply by changing the
learning rule. In this section, we discuss the problem of non-
stationary environments: How does the network maintain a
representation of stimuli that temporarily drop out from the
training set? We discuss how existing CLS models fail to solve
this problem, and how this problem can be addressed by adding
a new kind of off-line learning that rehearses and protects
existing knowledge structures.2
3.1. Why we need two kinds of off-line learning
The original form of the Complementary Learning Systems
framework as proposed by McClelland et al., 1995 included a
single form of off-line learning in which hippocampus replayed
memories to cortex. The role of this off-line learning was to
allow the cortical model to incorporate information about rare
events. As discussed below, this framework works well when
the environment is stationary (i.e. the composition of the
training set does not change) but it fails to preserve existing
knowledge when the environment is not stationary. This point
can be illustrated by considering what happens to our
knowledge of typical birds after seeing a penguin. We will
consider two situations: the stationary environment case
(where the subject continues to see typical birds) and the
non-stationary environment case (where typical birds are
temporarily removed from the environment).
Penguin learning with a stationary environment. In this
case, the person sees typical (winged, feathery, flying) birds on
a regular basis during waking. One day, the person goes to the
zoo and sees a (winged, feathery, flightless) penguin. The next
day, the person returns to seeing typical birds. The original
CLS model learns about penguins by taking a hippocampal
‘snapshot’ of the penguin, and then re-playing this memory to
cortex. Hippocampal playback of ‘penguins do not fly’ will
incrementally degrade the network’s knowledge that (typi-
cally) birds fly. However, if the network continues to encounter
typical birds (with high frequency) during waking, learning
about these typical birds will repair the damage done by off-
line learning about penguins.
Penguin learning with a non-stationary environment. In this
case, the person sees typical birds on a regular basis during
waking. Then, the person takes a month-long trip to Antarctica
during which they only see penguins (never typical birds). In
this case, hippocampal playback of new penguin memories and
repeated environmental exposure to penguins will degrade the
network’s knowledge about typical birds. Because (in this
example) typical birds are not present in Antarctica, learning-
during-waking will not help repair the network’s knowledge.
The only possible source of support for typical birds in this
situation is hippocampal replay of ‘typical bird’ memories
from before the trip. However, as time passes, new information
will be encountered in Antarctica that will also require off-line
playback. Gradually, the probability that pre-trip information
will be replayed, relative to Antarctica memories, will become
extremely small. Ultimately, when neither the environment nor
the hippocampus provides cortex with additional exposure to
pre-trip information, it will fade from cortex.3
3.2. Lessons from the penguin example
The penguin example illustrates that (in the original form of
CLS) the environment is responsible for repairing damage to
existing knowledge. When existing knowledge continues to be
reinforced by stimuli in the environment, CLS does fine. But, if
the environment changes (such that existing knowledge is no
longer directly supported by the environment) then the network
will show high levels of interference.
The fact that the network shows some forgetting of typical
birds, in and of itself, is not damning: From a computational
perspective, it is appropriate to decrease the prominence of
flying (vs. flightless) birds in semantic memory if the base rate
of encountering flying birds decreases. However, the excessive
speed of forgetting exhibited by the CLS model (and other
models like it) is highly problematic. Taken literally, this
property of the CLS model would imply that a person who
regularly spends summers in Antarctica and the rest of the year
in New Jersey would forget everything about New Jersey when
they go to Antarctica, and vice-versa. To address the problem
of catastrophic forgetting in non-stationary environments, we
suggest that a second off-line learning mechanism is needed.
The role of the second mechanism would be to slow the rate of
erosion of pre-existing memories. This mechanism needs to be
able to strengthen memories in situations where they are not
being supported by the environment. In the next section, we
discuss how learning during REM sleep may help to protect
2The simulation work described in this section was conducted as part of
Adler Perotte’s senior thesis research at Princeton University.
3The above argument is based on the idea that, as the person spends more
and more time in Antarctica, the ratio of Antarctica episodic memories to pre-
trip episodic memories will increase, resulting in proportionately less rehearsal
of pre-trip episodic memories. This prediction depends critically on the rules
that govern which memories get replayed by the hippocampus. It is possible (in
principle) that one could devise a clever algorithm for hippocampal replay that
continues to give privileged status to pre-trip memories. However, in practice,
we are not sure how this goal could be accomplished.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281220
memories, and we present a neural network model of this
3.3. Data on sleep and learning
The need for two distinct kinds of off-line learning
(hippocampal replay of new memories to cortex, and repair
of pre-existing memories, respectively) converges strongly
with recently acquired data on sleep and learning (for reviews,
see Gais & Born, 2004; Paller & Voss, 2004; Ribeiro &
Nicolelis, 2004; Stickgold, 1998; Walker & Stickgold, 2004).
These findings suggest that slow wave sleep (SWS) and REM
sleep contribute to learning in distinct ways: SWS may support
hippocampal replay of new memories to cortex, and REM may
support tuning of pre-existing cortical and hippocampal
representations. We briefly review the evidence for this linkage
Evidence linking SWS to hippocampal replay. The strongest
evidence for hippocampal replay during SWS comes from
electrophysiological studies that have examined the relation-
ship between hippocampal activity in SWS vs. waking. Several
studies have found that patterns of neural activity observed
during waking events reappear in subsequent periods of sleep,
and this replay occurs more frequently in SWS than in REM
(see, e.g. Wilson & McNaughton, 1994 for evidence of replay
in SWS; but see Louie and Wilson, 2001 for evidence that
some replay occurs during REM; see Ribeiro et al., 2004 for a
direct comparison showing more replay in SWS than REM).
Other studies have found that, during SWS, hippocampal
replay of memories is coherent with cortical reactivation (Qin,
McNaughton, Skaggs, & Barnes, 1997). Also, sharp wave-
ripple activity in hippocampus has been shown to be
temporally correlated to sleep spindle oscillations in cortex
during SWS (Siapas & Wilson, 1998; Sirota, Scicsvari, Huhl,
& Buzsaki, 2003). Finally, Hasselmo (1999) observed that
acetylcholine levels in the hippocampus are lower during SWS
vs. waking and REM. Hasselmo (1999) goes on to describe
how low acetylcholine levels should facilitate retrieval of
stored hippocampal memories (e.g. by increasing the relative
strength of CA3 recurrents). Although there is much work to be
done in specifying the exact nature of the hippocampo–cortical
interaction during SWS, these findings are broadly consistent
with the idea that (during SWS) hippocampus is teaching
cortex about recent events. For additional discussion of this
point, see Buzsaki (1998); Gais and Born (2004); Hasselmo
(1999), and Sejnowski and Destexhe (2000); for computational
models of this process, see Alvarez and Squire (1994); Meeter
and Murre (in press).
Evidence linking REM to neural plasticity. There is
extensive evidence, both direct and indirect, suggesting that
REM sleep plays an important role in neural plasticity. For
example, theta oscillations, which have been correlated to
human memory formation (e.g. Sederberg, Kahana, Howard,
Donner, & Madsen, 2003), are prevalent during REM sleep
(Winson, 1993). On a cellular level, Ribeiro and Nicolelis
(2004) show that transcriptional factors, associated with the
formation of memories during waking, are up-regulated during
REM. Also, recent studies have found behavioral evidence that
directly relates REM to learning on non-declarative memory
tasks. For example, Smith, Nixon, and Nader (2004) found that
the number and density of rapid eye movements (REMs)
increased, relative to a control group, after subjects performed
difficult novel tasks (mirror tracing and tower of Hanoi).
Additionally, the number of REMs correlated with the degree
of improved performance following sleep.
Importantly, while several studies have found evidence for
hippocampo–cortical interactions during SWS, there is much
less evidence for hippocampo–cortical synchrony during REM.
For example, while theta oscillations are more prevalent in
REM than SWS, these oscillations are not synchronized
between hippocampus and cortex (Cantero, Atienza, Stickgold,
Kahana, Madsen and Kocsis, 2003). These results suggest that
REM involves separate learning processes occurring within
cortex and hippocampus, as opposed to transfer of information
from hippocampus to cortex (for additional discussion of how
REM could tune cortical representations, see, e.g. Hasselmo,
3.4. The REM sleep model
In this section, we provide a brief overview of our model of
REM sleep. Based on the data reviewed above, it seems safe to
conclude that some kind of learning occurs during REM.
However, it is not clear (based on this data) how REM achieves
the functional goal of repairing damaged memories. The goal
of the modeling work presented here is to bridge the gap, and
show (to a first approximation) how a process with the
physiological properties described above can support memory
protection and repair.4
The most critical functional properties of REM, as reviewed
above, are: (1) cortical neural activity is unaffected by
environmental stimuli and uncorrelated with hippocampal
activity, and (2) theta oscillations are prevalent. As such, we
have modeled REM sleep as a period in which the cortical and
hippocampal networks are dissociated from external input (and
from each other) and autonomously rehearse stored memories.
In keeping with the finding of strong theta activity during
REM, learning during REM in our model is guided by
inhibitory oscillations (as per the Oscillating Learning
Algorithm section above).
The simulation of REM sleep presented here uses a cortical
memory architecture. We discuss later how to re-integrate the
hippocampal network into the model. With respect to cortical
learning, we view REM as a period where cortex can ‘think
about what it already knows’, thereby reinforcing knowledge
that may no longer be supported by the environment or by the
hippocampus. In our model, we initiate REM rehearsal by
presenting the cortical network with a single noisy input and
allowing the network to activate a memory. Once the REM
4Note that our model of the REM sleep process should not be confused with
the Shiffrin and Steyvers (1997) REMmodel of recognitionmemory,which has
nothing to do with REM sleep.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281221
episode is initiated, no further input is given to the network.
During REM rehearsal, the network transitions from one
attractor state to another due to synaptic depression, which
temporarily weakens the active pattern.
The model’s ability to repair damaged memories depends
critically on non-linear attractor dynamics in the cortical
network. These attractor dynamics give the network the ability
to recall the intact version of a memory even if the synapses
underlying that memory have been disrupted. There are clearly
limits to this dynamic: After a certain amount of damage, the
memory will simply cease to exist as an attractor state in the
cortical network. However, there is a relatively large window
where disrupting the underlying synaptic substrate of a
memory does not compromise recall of the pattern. This is
analogous to a building where the support beams are crumbling
but the building is still standing. If the REM rehearsal process
succeeds in finding a memory in this state (i.e. where it can still
be recalled, but the synaptic substrate is weak), then the
memory can be repaired.
In our model, learning during REM uses the same
oscillation-based learning algorithm that was used in the
pattern completion and familiarity simulations presented
earlier. As memories are rehearsed, the model oscillates the
strength of inhibition, and changes weights based on changes in
activation triggered by these inhibitory oscillations. We
discussed earlier how raising inhibition allows us to ‘stress-
test’ a memory. If a memory is weak (because of damage
incurred during SWS or awake learning, or because of
inadequate training), it will show decreased activity when we
raise inhibition, which, in turn, will trigger learning processes
that strengthen the memory. At an intuitive level, one can think
of SWS and awake learning as ‘denting’ existing memories
(like you would dent a car) but not destroying them; REM
learning provides a way of repairing these dents. The second
component of the learning algorithm (weakening memories
that activate when inhibition is lowered) also plays a critical
role. This competitor punishment mechanism allows existing
memories to push away new memories that are encroaching on
their space. More generally, this mechanism works to ensure
that memories retain their individuating features and do not all
collapse together (a problem that affects other models of
memory consolidation; see the Preventing runaway consolida-
tion section for additional discussion of this point). The idea
that competitor punishment occurs during sleep also leads to
specific behavioral predictions regarding effects of sleep on
memory, as discussed in the Applications to specific findings
3.5. Simulation: AB–AC interference
We used a simple list learning paradigm to explore how
incorporating REM affects learning and forgetting in our
cortical model. In particular, we were interested in exploring
how REM affects learning in situations where the environment
is not stationary. The network architecture consisted of three
50-unit layers (input, output, and hidden). For this set of
simulations, the input layer was bidirectionally connected to
the hidden layer, and the hidden layer was bidirectionally
connected to the output layer; there were no recurrent
connections within layers. Inhibition was oscillated on the
input and output layers. The training patterns for our model
were created in the spirit of the McCloskey and Cohen (1989)
AB–AC model of catastrophic forgetting. The ‘AB’ list
consisted of 15 randomly generated input–output pairs. The
‘AC’ list was generated by taking each pattern from the AB list
and changing three units (out of 5) from both the input and
output patterns such that each new pattern was 40% similar to
the corresponding AB pattern. This high level of similarity
between the two lists made it difficult for the network to
maintain memories of the AB items as it learned the AC items.
The network was initially trained on the AB items to
criterion. Next, we started training the network on AC patterns.
For these trials, the AC pattern was presented directly to the
network. These AC trials can be viewed as a proxy for learning
occurring during waking and SWS. In the REM Sleep
condition, the network was allowed to do REM rehearsal
after each epoch of AC training. In the No REM Sleep
condition, the network was not allowed to do any REM
rehearsal in between AC epochs. After each epoch of AC
training, we tested the network’s memory for all of the AB and
AC patterns by presenting the input-layer pattern and
measuring the network’s ability to recall the corresponding
output-layer pattern (note that learning was turned off during
test trials, and inhibition was not oscillated at test). By
comparing the REM Sleep and No REM Sleep conditions, we
were able to explore how including REM affects retention of
AB items and learning of AC items.
3.6. Implementation of REM
Basic REM parameters. The learning parameters used
during REM were identical to the learning parameters used
during AB and AC learning, except for the fact that we used a
smaller inhibitory oscillation size during REM vs. AB and AC
learning. The most important difference between REM vs. AB
and AC learning is that external patterns were not applied to the
network during REM. Each REM episode was started by
initializing the activity values of the network to a random
value. After this initial burst of noise, the network was allowed
to autonomously generate patterns to rehearse. Each REM
episode lasted for 30,000 time steps. This value was selected
because it allowed the network to visit and repair enough
representations during REM to stabilize memories from the AB
list. The oscillating learning algorithm was run continuously
through the REM episode. Weight change values were
computed on a time-step-by-time-step basis during the REM
episode, but the weight changes were not actually implemented
until the end of the REM episode.
Preventing runaway consolidation. An important problem
that autonomous rehearsal mechanisms need to solve is
runaway consolidation (Ans & Rousset, 2000; Meeter, 2003;
Wittenberg, Sullivan, & Tsien, 2002). This problem arises
when some memories are stronger than others. When random
noise is injected into the system, the probability of recalling a
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281222
memory is a function of that memory’s strength. Thus, strong
memories are rehearsed more often than weak memories. This
leads to a positive feedback loop: Patterns that are rehearsed
become even stronger, which makes them even more likely to
be rehearsed in the future. This pattern of rehearsal leads to a
situation where a small number of memories become extremely
strong, and all other memories in the system become extremely
weak. Rehearsal algorithms that manifest this problem are
obviously unsuited for the task of preserving stored
To prevent the network from repeatedly settling into the
same patterns, we implemented a synaptic depression
mechanism. This mechanism slowly reduces the efficacy of
connections between concurrently active units. As time
progresses, the active pattern tires and dissipates. When this
happens, other units activate and the network settles into a
new pattern. Once a depressed connection is no longer being
used, it begins to rebuild its strength. We selected a synaptic
depression mechanism that depends on both presynaptic and
postsynaptic activity, rather than a mechanism that depends
only on presynaptic activity (e.g. Huber & O’Reilly, 2003;
Gotts & Plaut, 2002), because the former mechanism is much
more specific in targeting the active memory. When
depression depends entirely on presynaptic activity, it will
generalize to all memories that share neurons with the active
memory, whereas depression that depends on presynaptic and
postsynaptic activity will only generalize to (the smaller set
of) memories that share synapses with the active memory.
Having said this, however, the basic pattern of results reported
in the next section does not depend on our use of the ‘postC
pre’ synaptic depression mechanism; the same qualitative
pattern was found when we used depression based on
presynaptic activity only.
3.7. Learning with and without REM
The inclusion of the REM episodes after each epoch of AC
training greatly reduced the rate of forgetting of the AB items.
This can be seen in Fig. 4. Without REM, the average number
of AB patterns recalled dropped below 2 (out of 15) after 20
epochs of AC training. In contrast, with the inclusion of REM,
the network was able to retain more than 11 of the AB items
after 20 epochs of AC training. In addition to reducing
forgetting of AB items, REM also slowed down acquisition of
AC items. The network was able to learn all 15 AC patterns
both with REM and without REM, but this process took
approximately 12 epochs of training with REM, vs. 6 epochs
without REM. The slower pace of learning with REM reflects
the fact that weaving new memories in with old memories
(without destroying the old memories) is a more demanding
process than simply letting the new memories overwrite the old
memories. In the former case, the network has to shuffle around
representations to make room for both AB and AC memories,
whereas in the latter case the network can simply re-use the
same set of neurons. Because REM model learned the same
number of AC memories and retained more AB memories, the
total number of patterns stored in the network at the end of
training was larger with REM than without REM (26 vs. 16).
3.8. REM discussion
In summary: Adding ‘REM sleep’ periods to the McClel-
land et al. (1995) Complementary Learning Systems model
significantly reduces the amount of forgetting. In this section,
we briefly review how our model relates to other theoretical
accounts of memory protection, and we discuss future
directions for the model.
3.9. Relation to other computational models of memory
Our model of how REM preserves memories can be viewed
as a descendant of models proposed by French (1997) and Ans
and Rousset (1997), and later by Ans and Rousset (2000).
Catastrophic Interference Without REM
Epochs of AC Training
02468 10 12 14 16 18 20
Memory Preservation With REM
Epochs of AC Training
02468 10 12 14 16 18 20
Fig. 4. Graphs of how AC training affects memory for AB and AC items, both with and without REM. Each graph plots the number of AB and AC items correctly
recalled, after each epoch of AC training. The left-hand graph shows the model’s performance without REM. The right-hand graph shows the model’s performance
with REM. A comparison of the two figures shows that including REM episodes greatly reduces the forgetting of the AB items, at the cost of slightly slowing
acquisition of AC items.
5To solve the problem of runaway consolidation, it is not necessary to
completely eliminate effects of memory strength on rehearsal. We suspect that
this is not possible, nor would it be desirable in light of behavioral data
suggesting that (during awake learning of word lists) strong items are rehearsed
more often than weak items (e.g. Ward, Woodward, Stevens, & Stinson, 2003).
Rather, the goal is to ensure that weak memories continue to be rehearsed, to a
degree that is sufficient to preserve these memories.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281223
These studies pioneered the use of random noise to elicit (and
then learn about) stored memory patterns (see also Wittenberg
et al., 2002). The main difference between our model of
cortical memory preservation and the Ans and Rousset (2000)
model is that Ans and Rousset (2000) use two networks (with
basically identical properties) to implement cortical memory
preservation, whereas our model uses a single, unified network.
The second network in their model maintains pristine copies of
older memories, which the first network can subsequently use
to repair representations that were damaged in new learning.
The primary contribution of our work is to show that damage
caused by new learning can be repaired without consulting an
undamaged copy of the knowledge base. As discussed above,
our scheme exploits the fact that—when synaptic weights have
been disrupted by a relatively small amount—it is still possible
to retrieve the memory in its original form. Thus, so long as the
damaged memory is located (during REM rehearsal) before it
becomes unrecallable, it can be repaired. This allows us to
dispense with the neurobiologically implausible ‘second
cortical network’ posited by Ans and Rousset (2000).
Other solutions to the stability–plasticity problem, such as
Carpenter and Grossberg’s Adaptive Resonance Theory (ART;
Carpenter & Grossberg, 1988, 2002, 2003), do not require off-
line learning. ART avoids catastrophic interference by gating
when learning occurs. The gating mechanism prevents learning
when the current pattern differs too much from top-down
expectations (and, thus, learning the current pattern would
significantly alter these top-down expectations). Although it is
impressive that ART does not require off-line learning, this
functional strength can also be viewed as an explanatory
shortcoming: Because ART does not need off-line learning, it
does not provide a natural explanation for data showing that
off-line learning (during sleep) actually occurs.
One other model of note is the cortico–hippocampal model
published by Kali and Dayan (2004). In this paper, Kali and
Dayan (2004) present simulations showing that hippocampal
replay alone (in the absence of an extra ‘memory protection’
process) can help to preserve semantic memories after the
statistics of the training environment change. On the surface,
this result appears to be inconsistent with our claim that an
extra memoryprotection process is requiredto fully address the
catastrophic interference problem. However, the utility of
the Kali and Dayan (2004) model (with regard to solving the
catastrophic interference problem) is compromised by two
issues. First, the simulations presented in Kali and Dayan
(2004) only explore the effect of subsequent cortical learning
on memory storage, not the effect of subsequent hippocampal
learning (specifically: no new hippocampal memories are
formed after the statistics of the training environment change).
Furthermore, the hippocampal component of the model is not
explicitly simulated, so they cannot explore the possibility that
new hippocampal memories might distort previously stored
hippocampal memories. We suspect that updating the Kali and
Dayan (2004) model to address these issues (by adding new
episodic learning after the training environment changes,
and allowing for the possibility of interference within
the hippocampus) would greatly compromise their model’s
ability to preserve stored semantic knowledge.
3.10. Future directions
In this section, we have provided a simple demonstration of
how adding a ‘REM sleep’ mechanism to CLS can help to
minimize interference. That said, there is a great deal of work
that remains to be done in understanding the neural
mechanisms that support off-line learning (and their functional
consequences). In this section, we describe ways in which the
model can be refined and extended, and ways in which the
model can be applied to specific sleep and learning findings.
Re-integrating the hippocampal model. Having demon-
strated the basic properties and feasibility of the REM memory
protection mechanism (as applied to cortex), the next logical
step is to add the hippocampal network from Norman and
O’Reilly (2003) back into the model. Re-integrating the
hippocampal model would allow us to explicitly model
waking, SWS, and REM sleep. During waking, the hippo-
campus would learn with a very high learning rate, and cortex
would learn with a much smaller learning rate (as per basic
CLS principles). During SWS, the hippocampal network would
recall memories acquired during the waking state through a
random settling process (similar to that used during REM in the
cortical model), and cortex would learn based on these
hippocampal training trials. As per the ideas described in
Hasselmo (1999), we would adjust modulatory parameters in
the hippocampal model to facilitate retrieval during SWS (vs.
encoding during waking and REM). The model would be
configured to stay in SWS long enough to sample recently
acquired hippocampal memories, but not so long that SWS
destroys the attractor network of the cortical model. After
SWS, we would implement the REM memory protection
process in both the cortex and the hippocampus independently.
Although we used a cortical architecture in the simulations
described above, the same basic principles of autonomous re-
activation and strengthening can also be applied to the
hippocampal model. As mentioned earlier, although there is
less overlap between memories in hippocampus vs. cortex,
there is still some overlap, which leads to interference. If
enough interference builds up, this could prevent the
hippocampus from fulfilling its job of conveying recent
memories to cortex during SWS. As such, the hippocampus
(like cortex) stands to benefit from the memory protection
mechanisms discussed in this section.
Targeting damaged memories. In the REM simulation
presented in this paper, the REM rehearsal process was able to
sample (almost) all of the AB traces because there were only 15
of these traces. In a brain-size network, with thousands (or
millions) of attractor states, this kind of exhaustive sampling is
not possible. In this situation, the REM rehearsal process needs
to be able to selectively sample the relatively small set of
memories that have suffered the most damage during SWS.
A major future direction for the REM model is to explore
mechanisms that will promote selective sampling of damaged
(vs. non-damaged) memories.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–1228 1224
There are a number of potential solutions to this problem.
One possibility is to incorporate a weak influence of the
hippocampus during the process of REM. According to this
view, the hippocampal network would continue to recall
memories, but (unlike SWS) this hippocampal influence would
not be strong enough to force a pattern of activity on the
cortical network. Rather, its influence on cortex would serve to
weakly guide activity to the areas of attractor space that were
visited during SWS (and, therefore, were most likely to have
been damaged). The weak hippocampal input would trigger re-
activation of cortical attractors in the ‘damaged’ areas, thereby
making it possible to repair these attractors. Another, related
possibility (which does not require hippocampo–cortical
interactions during REM) would be to apply a hysteresis
algorithm to units visited during SWS. This hysteresis
algorithm would enact a temporary increase in the efficacy of
neurons and/or synapses that were activated during SWS. If
hysteresis carried over (from SWS) into the following REM
phase, it would serve to guide the cortical network to regions of
attractor space that were visited during SWS.
One possible issue with both the ‘weak hippocampal
influence’ and ‘hysteresis’ ideas is that, by guiding cortex to
regions of attractor space visited during SWS, these mechan-
isms may foster additional strengthening of new memory
traces, as opposed to repair of old memory traces. However, we
do not think this is a major concern, for two reasons: First, if a
memory is truly new, then its cortical memory trace will be
weaker than the cortical memory traces of pre-existing
memories, so cortex will be more likely to rehearse the pre-
existing memories. Second, even if the network does rehearse
new memories (to some extent) during REM, eventually these
memories will tire out due to synaptic depression. If the
network stays focused on the same region of attractor space
(and new memories are depressed) then it will rehearse old
memories in that region.
Applications to specific findings. Another future direction is
to use the model to simulate specific sleep-and-learning
datasets. Over the past few years, several studies have been
published that go beyond proving the mere existence of
learning during sleep, and map out a more detailed landscape
of how sleep affects learning. In this section, we will sketch out
how our ideas about memory competition (and competitor
punishment) during REM can be applied to some puzzling data
from Walker, Brakefield, Hobson, and Stickgold (2003) on
how sleep affects memory for motor sequences.
The basic finding from this study is that sleep enhances
memory for simple motor sequences in a button-pressing task.
Walker et al. (2003) build on this finding in several different
ways. In one variant of this paradigm, subjects learned one
sequence (S1) and then learned a second sequence (S2)
immediately afterward. Memory for S2 (measured in terms of
accuracy) improved after sleep, but memory for S1 did not.
However, when six (waking) hours intervened between
learning S1 and S2, both sequences showed improved accuracy
after sleep. Walker et al. (2003) explain this finding in terms of
the idea that 6 h of waking can ‘stabilize’ a memory, thereby
protecting it from interference from subsequent learning.
However, this pattern of results can also be explained in
terms of competitive dynamics during REM. In the ‘no delay’
condition, the two memories are encoded in a very similar
spatiotemporal context. This contextual overlap makes the
memory traces associated with S1 and S2 more similar, which
in turn increases the extent to which the two memories compete
during sleep. Since S2 is stronger, it is more likely to win the
competition (and S1 is more likely to lose), which implies that
S2 will benefit more from REM than S1. In contrast, when 6 h
intervene between learning the sequences, the two sequences
will be associated with relatively distinct sets of contextual
features (e.g. you might be hungry when learning S1 but not
when learning S2). As a result, the cortical engrams of S1 and
S2 will be more different in this condition than in the ‘no delay’
condition. Because there is less overlap, the memory traces are
less likely to compete, so they both should benefit equally from
In a related finding, Walker et al. (2003) trained subjects on
S1 and let them sleep (so S1 performance improved). On the
second day of the experiment, Walker et al. (2003) trained
subjects on S2. Prior to learning S2, some subjects were briefly
re-exposed to S1, and some subjects were not. On the third day,
all subjects showed improved memory for the S2.However, the
re-exposure manipulation had a large effect on memory for S1:
Subjects who were re-exposed to S1 showed a large decrease in
S1 performance from day 2 to day 3. In contrast, subjects who
were not re-exposed to S1 did not show a change in S1
performance. Walker et al. (2003) interpret these results in
terms of reconsolidation; according to this idea, reactivating a
memory temporarily makes the molecular substrate of that
memory more labile, and thus more vulnerable to interference.
For example, in the animal literature, several studies have
found that ‘reminding’ an animal of a tone-shock association
(by presenting the tone by itself) makes that tone-shock
memory vulnerable to disruption via injection of a protein
synthesis inhibitor (e.g. Nader, Schafe, & LeDoux, 2000; for a
recent review, see Dudai & Eisenberg, 2004; for additional
discussion of how the Walker et al., 2003 finding relates to the
animal reconsolidation literature, see Nader, 2003). However,
we can explain these results without positing re-labialization of
the S1 memory. Rather, according to our REM framework,
presenting S1 in the same context as S2 makes it more likely
that S1 will pop up as a competitor to S2 during the second
night of sleep, which (in turn) will hurt memory for S1.
We still need to build a working simulation of the Walker et
al. (2003) data, in order to test the sufficiency of these ideas.
However, even without a working simulation, it is not difficult
to generate testable predictions that follow from our
competitor-punishment account of these results. According to
our theory, the key variable that determines whether S1
memory improves during sleep is the similarity of S1 to S2.
This implies that, if we could make the S1 memory trace more
different from the S2 trace (by changing the stimuli or motor
responses, or by having subjects learn the two sequences in
different rooms), this would allow the network to rehearse both
S1 and S2 without them interfering with one another. In
contrast, the reconsolidation account does not intrinsically
K.A. Norman et al. / Neural Networks 18 (2005) 1212–1228 1225
make predictions about effects of S1–S2 similarity (although it
is not incompatible with the idea that higher similarity will lead
to higher interference).
The fact that we can account for the Walker et al. (2003)
data in terms of competitor punishment has led us to consider
whether we can account for other reconsolidation findings in
terms of competitor punishment. Specifically, in the fear
conditioning paradigm described above (where subjects are
exposed to a tone-shock pair, and later are ‘reminded’ of this
association by presenting the tone), it is possible to construe the
‘tone alone’ presentation as a competitor to the original
memory (i.e. tone is now being paired with safety, not shock),
rather than a reminder of the original memory (see Eisenberg,
Kobilo, Berman, & Dudai, 2003 for a similar idea). Given this
premise, we can construct a competitor-punishment account of
basic reconsolidation findings:
†One of the key ideas presented in this paper is that
catastrophic interference is the ‘default mode’ for neural
networks: In the absence of special memory protection
mechanisms (e.g. the REM sleep mechanism presented here),
new learning will weaken similar pre-existing memories.
Rehearsal of the new memory during SWS will compound this
effect, resulting in worse and worse recall of the pre-existing
memory over time.
†In the fear conditioning paradigm, after the animal is
exposed to the ‘tone-safety’ association, we posit that REM
learning (or something like it) is needed to protect the original
†Protein synthesis blockers may disrupt REM learning (see
Ribeiro & Nicolelis, 2004 for discussion of transcriptional
factors in REM sleep) while leaving other forms of neural
plasticity relatively intact. In this situation (where learning
about the tone-safety event is still taking place, in the absence
of REM memory protection), the tone-safety memory will
catastrophically interfere with the tone-shock memory,
resulting in diminished fear conditioning.
At this point, these ideas are highly speculative (especially
the idea that protein synthesis blockers would selectively
disrupt REM memory protection mechanisms). However, in
light of recent attempts to explain human learning data in terms
of reconsolidation theory, we think it is equally important to
consider whether animal ‘reconsolidation’ findings can be
explained by network-level interference theories such as CLS.
At the very least, this leads one to consider factors that were
previously neglected (e.g. the exact relationship between the
‘reminder’ and the original memory; and also the role of
different sleep stages in promoting/preventing reconsolidation
3.11. Final thoughts
In recent years, Complementary Learning Systems research
has focused on simulating specific findings (e.g. Norman &
O’Reilly, 2003; O’Reilly & Rudy, 2001). While this approach
has been very productive, it is alsoimportant to take a step back
and assess how well CLS does at addressing the problem that it
was designed to solve: accurately storing and maintaining
knowledge about the environment. In this paper, we have
outlined the challenges that the CLS model faces, in order to
provide a satisfactory solution to the stability–plasticity
problem. We have also tried to outline some possible ways
of addressing these challenges. We showed how leveraging
inhibitory oscillations can help reduce interference during
learning, by ensuring that synapses are modified judiciously
(i.e. such that strengthening is focused on weak target units,
and weakening is focused on strong non-target units). We also
showed how a ‘REM sleep’ process can be used to protect
memories when these memories are no longer being directly
supported by the environment. Importantly, although our
discussion of the oscillating learning algorithm and REM
sleep has focused on the cortical model, we think that the
oscillating algorithm and the REM memory protection
mechanism may also be applicable to the hippocampus. We
are presently exploring this possibility.
The analyses presented here illustrate the vast distance that
CLS has to travel before it solves stability–plasticity (e.g. we
need to devise a mechanism that will allow the REM sleep
model to scale up to larger networks, with more stored
memories). However, CLS has a long history of turning its
weaknesses into strengths: Understanding that networks with
distributed, overlapping representations perform poorly at
rapid memorization led to a better understanding of why we
need a hippocampus, and how it works. Understanding that the
standard CLS model fails to deal properly with non-stationary
environments may lead to a better understanding of REM
sleep. Finally, understanding the limitations of simple Hebbian
learning may lead to a better understanding of the functional
role of theta oscillations. This history gives us hope that—as
we continue to chip away at stability–plasticity—our efforts
will be repaid with further insights into the functional and
neural architecture of learning.
Ackley, D. H., Hinton, G. E., & Sejnowski, T. J. (1985). A learning algorithm
for Boltzmann machines. Cognitive Science, 9, 147–169.
Aggleton, J. P., & Brown, M. W. (1999). Episodic memory, amnesia, and the
hippocampal-anterior thalamic axis. Behavioral and Brain Sciences, 22,
Alvarez, P., & Squire, L. R. (1994). Memory consolidation and the medial
temporal lobe: A simple network model. Proceedings of the National
Academy of Sciences, USA, 91, 7041–7045.
Anderson, M. C. (2003). Rethinking interference theory: Executive control and
the mechanisms of forgetting. Journal of Memory and Language, 49, 415–
Anderson, M. C., & Bell, T. (2001). Forgetting our facts: the role of inhibitory
processes in the loss of propositional knowledge. Journal of Experimental
Psychology: General, 130(3), 544–570.
Ans, B., & Rousset, S. (1997). Avoiding catastrophic forgetting by coupling
two reverberating neural networks. Academie des Sciences, Sciences de la
vie, 320, 989–997.
Ans, B., & Rousset, S. (2000). Neural networks with a self-refreshing memory:
knowledge transfer in sequential learning tasks without catastrophic
forgetting. Connection Science, 12(1), 1–19.
Blaxton, T. A., & Neely, J. H. (1983). Inhibition from semantically related
primes: Evidence of a category-specific retrieval inhibition. Memory and
Cognition, 11, 500–510.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–1228 1226
Bogacz, R., & Brown, M. W. (2003). Comparison of computational models of
familiarity discrimination in the perirhinal cortex. Hippocampus, 13, 494–
Burgess, N., & O’Keefe, J. (1996). Neuronal computations underlying the
firing of place cells and their role in navigation. Hippocampus, 6, 749–762.
Buzsaki, G. (1998). Memory consolidation during sleep: A neurophysiological
perspective. Journal of Sleep Research, 7, 17–23.
Buzsaki, G. (2002). Theta oscillations in the hippocampus. Neuron, 33, 325–
Cantero, J. L., Atienza, M., Stickgold, R., Kahana, M. J., Madsen, J. R., &
Kocsis, B. (2003). Sleep-dependent theta oscillations in the human
hippocampus and neocortex. Journal of Neuroscience, 23, 10897–10903.
Carpenter, G. A., & Grossberg, S. (1988). The art of adaptive pattern
recognition by a self-organizing neural network. Computer, 21(3), 77–88.
Carpenter, G. A., & Grossberg, S. (2002). A self-organizing neural network for
supervised learning, recognition, and prediction. In T. A. Polk, & C. M.
Seifert (Eds.), Cognitive modeling (pp. 289–314). Cambridge, MA: MIT
Carpenter, G. A., & Grossberg, S. (2003). Adaptive resonance theory. The
handbook of brain theory and neural networks (pp. 87–90). Cambridge,
MA: MIT Press.
Ciranni, M. A., & Shimamura, A. P. (1999). Retrieval-induced forgetting in
episodic memory. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 25, 1403.
Dudai, Y., & Eisenberg, M. (2004). Rites of passage of the engram:
Reconsolidation and the lingering consolidation hypothesis. Neuron, 44,
Eichenbaum, H., Otto, T., & Cohen, N. J. (1994). Two functional components
of the hippocampal memory system. Behavioral and Brain Sciences, 17(3),
Eisenberg, M., Kobilo, T., Berman, D. E., & Dudai, Y. (2003). Stability of
retrieved memory: Inverse correlation with trace dominance. Science, 301,
French, R. M. (1997). Pseudo-recurrent connectionist networks: An approach
to the ‘sensitivity-stability’ dilemma. Connection Science, 9, 353–379.
French, R. M. (1999). Catastrophic forgetting in connectionist networks:
Causes, consequences and solutions. Trends in Cognitive Sciences, 3(4),
French, R. M. (2003). Castrophic forgetting in connectionist networks. In L.
Nadel (Ed.), Encyclopedia of cognitive science. London: MacMillan.
Gais, S., & Born, J. (2004). Declarative memory consolidation: Mechanisms
acting during human sleep. Learning and Memory, 11, 679–685.
Gotts, S. J., & Plaut, D. C. (2002). The impact of synaptic depression following
brain damage: A connectionist account of ‘access/refractory’ and
‘degraded-store’ semantic impairments. Cognitive, Affective, and Beha-
vioral Neuroscience, 2(3), 187–213.
Grossberg, S. (1976). Adaptive pattern classification and universal recoding. I.
Parallel development and coding of neural feature detectors. Biological
Cybernetics, 23, 121–134.
Hasselmo, M. E. (1999). Neuromodulation: Acetylcholine and memory
consolidation. Trends in Cognitive Sciences, 3(9), 351–359.
Hasselmo, M. E., Bodelon, C., & Wyble, B. P. (2002). A proposed function for
hippocampal theta rhythm: Separate phases of encoding and retrieval
enhance reversal of prior learning. Neural Computation, 14, 793–818.
Hasselmo, M. E., & Wyble, B. (1997). Free recall and recognition in a network
model of the hippocampus: Simulating effects of scopolamine on human
memory function. Behavioural Brain Research, 89, 1–34.
Hinton, G. E. (1989). Deterministic Boltzmann learning performs steepest
descent in weight-space. Neural Computation, 1, 143–150.
Hinton, G. E., & McClelland, J. L. (1988). Learning representations by
recirculation. In D. Z. Anderson (Ed.), Neural Information Processing
Systems, 1987, 358–366.
Hinton,G. E., & Sejnowski, T. J. (1986).Learning and relearning in Boltzmann
machines. In D. E. Rumelhart, J. L. McClelland, & PDP Research Group,
Foundations. Parallel distributed processing (Vol. 1) (pp. 282–317).
Cambridge, MA: MIT Press.
Holscher, C., Anwyl, R., & Rowan, M. J. (1997). Stimulation on the positive
phase of hippocampal theta rhythm induces long-term potentiation that can
be depotentiated by stimulation on the negative phase in area CA1 in vivo.
Journal of Neuroscience, 17, 6470.
Huber, D. E., & O’Reilly, R. C. (2003). Persistence and accommodation in
short-term priming and other perceptual paradigms: temporal segregation
through synaptic depression. Cognitive Science, 27, 403–430.
Huerta, P. T., & Lisman, J. E. (1996). Synaptic plasticity during the cholinergic
theta-frequency oscillation in vitro. Hippocampus, 49, 58–61.
Hyman,J.M., Wyble,B. P.,Goyal,V.,Rossi, C. A., & Hasselmo, M. E.(2003).
Stimulation in hippocampal region CA1 in behaving rats yields long-term
potentiation when delivered to the peak of theta and long-term depression
when delivered to the trough. Journal of Neuroscience, 23, 11725–11731.
Kahana, M. J. (2001). Theta returns. Current Opinion in Neurobiology, 11,
Kali, S., & Dayan, P. (2004). Off-line replay maintains declarative memories in
a model of hippocampal-neocortical interactions. Nature Neuroscience, 7,
Koutstaal, W., Schacter, D. L., & Jackson, E. M. (1999). Perceptually based
false recognition of novel objects in amnesia: Effects of category size and
similarity to category prototypes. Cognitive Neuropsychology, 16, 317.
Louie, K., & Wilson, M. (2001). Temporally structured replay of awake
hippocampal ensemble activity during rapid eye movement sleep. Neuron,
Marr, D. (1971). Simple memory: A theory for archicortex. Philosophical
Transactions of the Royal Society (London) B, 262, 23–81.
McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are
complementary learning systems in the hippocampus and neocortex:
Insights from the successes and failures of connectionist models of learning
and memory. Psychological Review, 102, 419–457.
McCloskey, M., & Cohen, N. J. (1989). Catastrophic interference in
connectionist networks. The sequential learning problem. In G. H. Bowes,
The psychology of learning and motivation (Vol. 24) (pp. 109–164). San
Diego, CA: Academic Press.
McNaughton, B. L., & Morris, R. G. M. (1987). Hippocampal synaptic
enhancement and information storage within a distributed memory system.
Trends in Neurosciences, 10(10), 408–415.
Meeter, M. (2003). Control of consolidation in neural networks; avoiding
runaway effects. Connection Science, 15(1), 45–61.
Meeter, M., & Murre, J. (in press). Tracelink: A model of amnesia and
consolidation. Cognitive Neuropsychology.
Minai, A. A., & Levy, W. B. (1994). Setting the activity level in sparse random
networks. Neural Computation, 6, 85–99.
Moll, M., & Miikkulainen, R. (1997). Convergence-zone episodic memory:
Analysis and simulations. Neural Networks, 10, 1017–1036.
Movellan, J.R.(1990).Contrastive Hebbianlearningin thecontinuousHopfield
model. In D. S. Tourtezky, G. E. Hinton, & T. J. Sejnowski (Eds.),
Proceedings of the 1989 connectionist models summer school (pp. 10-17).
Nader, K. (2003). Re-recording human memories. Nature, 425, 571–572.
Nader, K., Schafe, G. E., & LeDoux, J. E. (2000). Fear memories require
proteinsynthesisin the amygdalaforreconsolidation after retrieval. Nature,
Norman, K.A., Newman, E.L., & Detre, G.J. (2004). Further predictions from a
neural network model of retrieval-induced forgetting. 45th Annual Meeting
of the Psychonomic Society. Minneapolis, MN.
Norman, K. A., Newman, E. L., Detre, G. J., & Polyn, S. M. (2005). How
inhibitory oscillations can train neural networks and punish competitors.
(Technical Report 05-1). Princeton, NJ: Princeton University, Center for
the Study of Brain, Mind, and Behavior.
Norman, K. A., & O’Reilly, R. C. (2003). Modeling hippocampal and
neocortical contributions to recognition memory: A complementary-
learning-systems approach. Psychological Review, 4, 611–646.
O’Keefe, J., & Nadel, L. (1978). The hippocampus as a cognitive map. Oxford,
England: Oxford University Press.
O’Reilly, R. C., & Munakata, Y. (2000). Computational explorations in
cognitive neuroscience: Understanding the mind by simulating the brain.
Cambridge, MA: MIT Press.
K.A. Norman et al. / Neural Networks 18 (2005) 1212–12281227
O’Reilly, R. C., & Norman, K. A. (2002). Hippocampal and neocortical Download full-text
contributions to memory: Advances in the complementary learning systems
framework. Trends in Cognitive Sciences, 12, 505–510.
O’Reilly, R. C., & Rudy, J. W. (2001). Conjunctive representations in learning
and memory: Principles of cortical and hippocampal function. Psychologi-
cal Review, 108, 311–345.
Paller, K. A., & Voss, J. K. (2004). Memory reactivation and consolidation
during sleep. Learning and Memory, 11, 664–670.
reprocessing in corticocortical and hippocampal neuronal ensembles.
Philosophical Transactions: Biological Sciences, 352, 1525–1533.
Raghavachari, S., Kahana, M. J., Rizzuto, D. S., Caplan, J. B., Kirschen, M. P.,
Bourgeois, B., et al. (2001). Gating of human theta oscillations by a
working memory task. Journal of Neuroscience, 9, 3175–3183.
Ribeiro, S., Gervasoni, D., Soares, E. S., Zhou, Y., Lin, S. C., Pantoja, J., et al.
(2004). Long-lasting novelty-induced neuronal reverberation during slow-
wave sleep in multiple forebrain areas. PLoS Biology, 2, 126–137.
Ribeiro, S., & Nicolelis, M. A. L. (2004). Reverberation, storage, and
postsynaptic propagation of memories during sleep. Learning and Memory,
Rolls, E. T. (1989). Functions of neuronal networks in the hippocampus and
neocortex in memory. In J. H. Byrne, & W. O. Berry (Eds.), Neural models
of plasticity: Experimental and theoretical approaches (pp. 240–265). San
Diego, CA: Academic Press.
Scoville, W. B., & Milner, B. (1957). Loss of recent memory after bilateral
hippocampal lesions. Journal of Neurology, Neurosurgery, and Psychiatry,
Sederberg, P., Kahana, M. J., Howard, M. W., Donner, E. J., & Madsen, J. R.
(2003). Theta and gamma oscillations during encoding predict subsequent
recall. Journal of Neuroscience, 23, 10809–10814.
Sejnowski, T. J., & Destexhe, A. (2000). Why do we sleep? Brain Research,
Sherry, D. F., & Schacter, D. L. (1987). The evolution of multiple memory
systems. Psychological Review, 94(4), 439–454.
Shiffrin, R. M., & Steyvers, M. (1997). A model for recognition memory: REM
— retrieving effectively from memory. Psychonomic Bulletin and Review,
Siapas, A. G., & &Wilson, M. A. (1998). Coordinated interactions between
hippocampal ripples and cortical spindles during slow-wave sleep. Neuron,
Sirota, A., Scicsvari, J., Huhl, D., & Buzsaki, G. (2003). Communication
betweenneocortexand hippocampus duringsleepin rodents.Neuroscience,
Smith,C. T., Nixon,M. R., & Nader,R. S. (2004).Posttrainingincreases in rem
sleep intensity implicate rem sleep in memory processing and provide a
biological marker of learning potential. Learning and Memory, 11, 714–
Squire, L. R. (1992). Memory and the hippocampus: A synthesis from findings
with rats, monkeys, and humans. Psychological Review, 99, 195–231.
Standing, L. (1973). Learning 10,000 pictures. Quarterly Journal of
Experimental Psychology, 25, 207–222.
Stickgold, R. (1998).Sleep: Off-line memory reprocessing. Trends in Cognitive
Sciences, 2, 484–492.
Sutherland, R. J., & Rudy, J. W. (1989). Configural association theory: The role
of the hippocampal formation in learning, memory, and amnesia.
Psychobiology, 17(2), 129–144.
Teyler, T. J., & Discenna, P. (1986). The hippocampal memory indexing
theory. Behavioral Neuroscience, 100, 147–154.
Toth, K., Freund, T. F., & Miles, R. (1997). Disinhibition of rat hippocampal
pyramidal cells by GABAergic afferents from the septum. Journal of
Physiology, 500, 463–474.
Treves, A., & Rolls, E. T. (1994). A computational analysis of the role of the
hippocampus in memory. Hippocampus, 4, 374–392.
Walker, M. P., Brakefield, T., Hobson, J. A., & Stickgold, R. (2003).
Dissociable stages of human memory consolidation and reconsolidation.
Nature, 425, 616–620.
Walker, M. P., & Stickgold, R. (2004). Sleep-dependent learning and memory
consolidation. Neuron, 44, 121–133.
Ward, G., Woodward, G., Stevens, A., & Stinson, C. (2003). Using overt
rehearsals to explain the word frequency effects in free recall. Journal
of Experimental Psychology: Learning, Memory, and Cognition, 29,
Wilson, M. A., & McNaughton, B. L. (1994). Reactivation of hippocampal
ensemble memories during sleep. Science, 265, 676–678.
Winson, J. (1993). The biology and function of rapid eye movement sleep.
Current Opinion in Neurobiology, 3, 243–248.
Wittenberg, G. M., Sullivan, M. R., & Tsien, J. Z. (2002). Synaptic reentry
reinforcement based network model for long-term memory consolidation.
Hippocampus, 12, 637–647.
Wu, X., Baxter, R. A., & Levy, W. B. (1996). Context codes and the effect of
noisy learning on a simplified hippocampal CA3 model. Biological
Cybernetics, 74, 159–165.
Yonelinas, A. P. (2002). The nature of recollection and familiarity: A
review of 30 years of research. Journal of Memory and Language, 46,
K.A. Norman et al. / Neural Networks 18 (2005) 1212–1228 1228