Amit's “Attractor Neural Network” perspective on cognition raises difficult technical problems already met by prior dynamical models. This commentary sketches briefly some of them concerning the internal topological structure of attractors, the constituency problem, the possibility of activating simultaneously several attractors, and the different kinds of dynamical structures one can use to model brain activity: point attractors, strange attractors, synchronized arrays of oscillators, synfire chains, and so forth.

Persistent activity can be the product of mechanisms other than attractor reverberations. The single-unit data presented by Amit cannot discriminate between the different mechanisms. In fact, single-unit data do not appear to be adequate for testing neural network models.

Interpreting the Miyashita et al. experiments in terms of a cellassembly representation does not adequately explain the performance of Miyashita's monkeys on novel stimuli. We will argue that the latter observations point to acompositionalrepresentation and suggest a dynamics involving rapid and reversible binding of distinct activity patterns.

This article addresses the ability of Parallel Distributed Processing (PDP) networks to generate stagewise cognitive development in accordance with Piaget's theory of cognitive epigenesis. We carried out a replication study of the simulation experiments by McClelland (1989) and McClelland and Jenkins (1991) in which a PDP network learns to solve balance scale problems. In objective tests motivated from catastrophe theory, a mathematical theory of transitions in epigenetical systems, no evidence for stage transitions in network performance was found. It is concluded (...) that PDP networks lack the ability to recover the positive outcomes of analogous catastrophe analyses of real cognitive developmental data. In an attempt to further characterize the learning behaviour of PDP networks, we carried out a second simulation study using the discrimination‐shift paradigm. The results thus obtained indicate that PDP learning is compatible with the learning of stimulus‐response relationships, not with the acquisition of mediating rules such as conceived in (neo‐)Piagetian theory. In closing, we speculate about the feasibility of simulating stagewise development with alternative network architectures. (shrink)

Hebbian cell-assembly theory and attractor networks are good starting points for modeling cortical processing. Detailed cell models can be useful in understanding the dynamics of attractor networks. Cell assemblies are likely to be distributed, with the cortical column as the local processing unit. Synaptic memory may be dominant in all but the first couple of seconds.

A distinction should be made between the formation of stimulus-driven associations and cognitive concepts. To test the learning mode of a neural network, we propose a simple and classic input-output test: the discrimination shift task. Feed-forward PDP models appear to form stimulus-driven associations. A Hopfield network should be extended to apply the test.

Further tests of Amit's model are indicated. One strategy is to use the apparent coding sparseness of the model to make predictions about coding sparseness in Miyashita's network. A second approach is to use memory overload to induce false positive responses in modules and biological systems. In closing, the importance of temporal coding and timing requirements in developing biologically plausible attractor networks is mentioned.

The concept of an attractor in a mathematical dynamical system is reviewed. Emphasis is placed on the distinction between a cell assembly, the corresponding attractor, and the attractor dynamics. The biological significance of these entities is discussed, especially the question of whether the representation of the stimulus requires the full attractor dynamics, or merely the cell assembly as a set of reverberating neurons. Comparison is made to Freeman's study of dynamic patterns in olfaction.

The cell assembly as a simple attractor cannot explain many cognitive phenomena. It must be a highly structured network that can sustain highly structured excitation patterns. Moreover, a cell assembly must be more widely distributed in space than on a square millimeter.

A theory of representation is incomplete if it states “representations areX” whereXcan be symbols, cell assemblies, functional states, or the flock of birds fromTheaetetus, without explaining the nature of the link between the universe ofXs and the world. Amit's thesis, equating representations with reverberations in Hebbian cell assemblies, will only be considered a solution to the problem of representation when it is complemented by a theory of how a reverberation in the brain can be a representation of anything.

Recurrent excitation is experimentally well documented in cortical populations. It provides for intracortical excitatory biases that linearize negative feedback interactions and induce macroscopic state transitions during perception. The concept of the local neighborhood should be expanded to spatial patterns as the basis for perception, in which large areas of cortex are bound into cooperative behavior with near-silent columns as important as active columns revealed by unit recording.

Amit argues that the local neuronal spike rate that persists (reverberating) in the absence of the eliciting stimulus represents the code of the eliciting stimulus. Based on the general argument that the inferred functional meaning of reverberation depends in part on the type of representational assumptions, reverberations may only be important for the encoding of contextual information.

Hebbian mechanisms are justified according to their functional utility in an evolutionary sense. The selective advantage of correlating content-contingent stimuli reflects the putative common cause of temporally or spatially contiguous inputs. The selective consequences of such correlations are discussed by using examples from the evolution of signal form in sexual selection and model-mimic coevolution. We suggest that evolutionary justification might be considered in addition to neurophysiology plansibility when constructing representational models.

This article addresses the ability of Parallel Distributed Processing (PDP) networks to generate stagewise cognitive development in accordance with Piaget's theory of cognitive epigenesis. We carried out a replication study of the simulation experiments by McClelland (1989) and McClelland and Jenkins (1991) in which a PDP network learns to solve balance scale problems. In objective tests motivated from catastrophe theory, a mathematical theory of transitions in epigenetical systems, no evidence for stage transitions in network performance was found. It is concluded (...) that PDP networks lack the ability to recover the positive outcomes of analogous catastrophe analyses of real cognitive developmental data. In an attempt to further characterize the learning behaviour of PDP networks, we carried out a second simulation study using the discrimination‐shift paradigm. The results thus obtained indicate that PDP learning is compatible with the learning of stimulus‐response relationships, not with the acquisition of mediating rules such as conceived in (neo‐)Piagetian theory. In closing, we speculate about the feasibility of simulating stagewise development with alternative network architectures. (shrink)

The cell assembly is an important concept for cognitive psychology. Cognitive processing will to a large extent depend on the relations that can exist between different assemblies. A potential relation between assemblies can already be seen in the occurrence of (classical) conditioning. However, the resulting associations between assemblies only produce behavioristic processing or so-called regular computation. Higher-level cognitive abilities most likely result from nonregular computation. I discuss the possibility of this form of computation in terms of cell assemblies.

The neurophysiological evidence from the Miyashita group's experiments on monkeys as well as cognitive experience common to us all suggests that local neuronal spike rate distributions might persist in the absence of their eliciting stimulus. In Hebb's cell-assembly theory, learning dynamics stabilize such self-maintaining reverberations. Quasi-quantitive modeling of the experimental data on internal representations in association-cortex modules identifies the reverberations (delay spike activity) as the internal code (representation). This leads to cognitive and neurophysiological predictions, many following directly from the language (...) used to describe the activity in the experimental delay period, others from the details of how the model captures the properties of the internal representations. (shrink)

Amit's efforts to provide stronger theoretical and empirical support for Hebb's cell-assembly concept is admirable, but we have serious reservations about the perspective presented in the target article. For Hebb, the cell assembly was a building block; by contrast, the framework proposed here eschews the need to fit the assembly into a broader picture of its function.

Every immediate memory is unique; it is therefore unlikely to consist of an attractor or even a combination of attractors. In the present state of knowledge about the chemistry of synaptic transmission, there is no reason to look beyond neurons that directly receive sensory afferents for the afterdischarges that correspond to active memories.

Amit's Hebbian model risks being overexplanatory, since it does not depend on specific physiological modelling of cortical ANNs, but concentrates on those phenomena which are modelled by a large class of ANNs. While offering a strong demonstration of the presence of Hebb's “cell assemblies,” it does not offer an equal account of Hebb's “phase sequence” concept.

Cortical reverberations may induce synaptic changes that underlie developmental plasticity as well as long-term memory. They may be especially important for the consolidation of synaptic changes. Reverberations in cortical networks should have particular significance during development, when large numbers of new representations are formed. This includes the formation of representations across different sensory modalities.

Amit defines cell assemblies aslocal cortical neuron populationswith strong internal connections. However, Hebb himself proposed that cell assemblies are distributed over different cortical areas (nonlocal ortranscortical assemblies). We review evidence from cognitive neuroscience and neuropsychology supporting the assumption that cell assemblies are transcortical.

The Miyashita experiments are very interesting and the results should be examined from a viewpoint of attractor dynamics. Amit's target article shows a path toward realistic modeling by artificial neural networks (ANN), but it is not necessarily the only one. I introduce another model that can explain a substantial part of the empirical observations and makes an interesting prediction. This model consists of such units that have nonmonotonic input-output characteristics with local inhibition neurons.

This commentary questions the target articles inferences from a limited set of empirical data to support this model and conceptual scheme. Especially questionable is the attribution of internal representation properties to an assembly of cells in a discrete cortical module firing at a discrete attractor frequency. Alternative inferences are drawn from cortical cooling and cell-firing data that point to the internal representation as a broad and specific cortical network defined by cortico-cortical connectivity. Active memory, it is proposed, consists in the (...) sustained activation of the component neuron populations of the network. (shrink)

The context of the target article is delimited again, underlining the intended locationof the argument in the bottomup hierarchy of brain study. The central message is that collective delay activity distributions (reverberations) in cortical modules extend the role of a spike (a potentialinformation carrier across long distances) to an active memory of structured, learned information that can be carried across long time intervals. Moreover, the population code of the reverberations makes them readable down the cortical processing stream. Most of the (...) critical comments are then interpreted and addressed in relation to misreading of the proper context. The price for the limitation of the context (in cognitive, behavioral, and computational terms) is compared with the advantages of a clear, direct contact with experiment on the one hand and with a well controlled body of modeling and analysis on the other. (shrink)