A central strength of the book is its detailed development and application of a single class of models. The book is recommended to a wide audience interested in how a particular neural network model is developed and then applied to a broad range of problems. Modular networks with locally distributed features, such as those present in CALM and other similar models, undoubtedly have important properties that are pertinent to biological and computational learning. Their relevance to psychology is being established. Ongoing investigations will hopefully provide more insight into how these interesting models may help us better understand brain processes and behavior.
1. In his book, Murre (1992) sets out an ambitious program of study dealing with several important notions in neurobiology, computation and psychology. The work focuses on modularity and learning in the context of local processing and noise-driven dynamics. A central strength of the book is its detailed development and application of a single class of models. This is in contrast to many recent books on neural networks that present overviews and updates on a variety of models, but pay little attention to how they are interrelated.
2. The categorization and learning module, CALM, contains subtypes of nodes that are linked in a distributed or connectionist manner. Behavioral features of a module are both locally autonomous and globally cooperative, once the module is placed within a larger network. These attributes are shared by modular systems in general and are beneficial in solving many tasks. In the case of CALM, Murre shows that these tasks include letter and pattern recognition, simulations of memory and memory loss and problems in genetics and evolution. Several well written appendices provide details of hardware and software implementations of CALM.
3. Although the book provides a comprehensive and readable account of an interesting model, the reader should be aware of several issues of biological, computational and psychological feasibility. These issues are outlined here in the context of determining what has been learned from the present study of modular neural networks.
4. Modularity is a principle of neuroscience that is present at multiple levels of structural and functional organization. To ascribe features of biological modularity to a model requires that enough detail be built into it to support the claim of neurobiological relevance. The heterogeneity of nodes, their complex wiring patterns, the arousal system, the modulation of learning and external inputs -- these are all characteristics that reflect biological motivation of the model. However, it is not obvious that CALM is a representation of the neocortical minicolumn based on its generic architecture and nonphysiological parameters.
5. Although any model will fall short in depicting a real system, it is unfortunate that CALM is not developed in a more biologically plausible manner. Two simple suggestions follow. First, refinement of the arousal and learning mechanisms in the model could be consolidated into a neuromodulatory system. This would operate in much the same way as the aminergic and cholinergic modulators, which arise from the pontine brainstem and basal forebrain and bathe the cortical modules, in a time varying manner (Hobson and Steriade, 1986). Second, wiring diagrams among the excitatory and inhibitory nodes within individual modules could be made more realistic (as suggested by the author's reference to Szentagothai, 1975).
6. Given that CALM is motivated by biology but is not inherently a model of biological computation, what principles of information processing are gathered from Murre's investigations? It is shown that within a module, prenodal activation patterns converge onto representational (R) nodes that interact with veto (V) and arousal (A) nodes. The convergence process is unique, but the associated mechanisms of learning, categorizing and separating correlated patterns are rather standard.
7. More interesting properties arise when several CALMs are linked together. It is here that the computational modeling shows the greatest promise since "modular architectures... stand somewhere between purely localized and fully distributed representations" (Murre, 1992, p. 127). The book lays some foundations and further work is apparently in progress, but the local and modular features of CALM, as a component of larger networks, have not been fully realized. This is a general comment. It is applicable to other modular architectures as well, including hierarchical cluster models (for example, Sutton et al., 1988), and emphasizes the point that novel computational and learning paradigms likely exist in modular yet distributed neural systems.
8. There is considerable enthusiasm about using artificial neural networks to probe mechanisms of normal and altered cognition (for example, see Farah, in press). In general, the models deal with functional networks, wherein the modules have behavioral but not necessarily structural significance. The behaviors examined by Murre include letter recognition, implicit and explicit memory, and catastrophic sequential interference. In common with other models of layered modules, such as the multi-layered perceptron and neocognitron, small networks of CALM modules produce results that are consistent with known psychological data.
9. Nevertheless, the predictive and explanatory properties of CALM are basically absent from Murre's discussion of CALM's psychological emulations. Without such information, it is difficult to assess the importance of CALM. Other models produce similar results, and this may be a reflection of fundamental principles of modular yet distributed networks rather than of CALM per se.
10. In summary, the book is recommended to a wide audience interested in how a particular neural network model is developed and then applied to a broad range of problems. Modular networks with locally distributed features, such as those present in CALM and other similar models, undoubtedly have important properties that are pertinent to biological and computational learning. Their relevance to psychology is being established. Ongoing investigations will hopefully provide more insight into how these interesting models may help us better understand brain processes and behavior.
Farah, M.J. Neuropsychological inference with an interactive brain: A critique of the locality assumption. Behavioral and Brain Sciences. In press.
Hobson, J.A., Steriade M. (1986) Neuronal basis of behavioral state control. In: Mountcastle V.B. (ed) Handbook of Physiology - The Nervous System, Vol IV. American Physiological Society: Bethesda, 701-823.
Murre, J.M.J. (1992) Learning and Categorization in Modular Neural Networks. UK: Harvester/Wheatsheaf; US: Erlbaum.
Murre, J.M.J. (1992) Precis of: Learning and Categorization in Modular Neural Networks. PSYCOLOQUY 3(68) categorization.1.
Sutton J.P., Beis J.S., Trainor L.E.H. (1988) Hierarchical model of memory and memory loss. Journal of Physics A: Math Gen 21, 4443-4454.
Szentagothai J. (1975) The "module concept" in cerebral cortex architecture. Brain Research 95, 475-496.