Murre introduces and discusses in detail a biologically and psychologically plausible neural network with impressive results. His model, however, has some distance to go before one could rightly claim that it is truly connected to either biology or psychology. Despite it's shortcomings I recommend this book for those who wish to be up-to-date on the latest in connectionist technique, if you are looking for a model of how cortex might actually work or of how we might actually think, then your expectations are too high. Murre's CALM is but a step in either of those directions.
1.1 Jacob Murre's (1992) book Learning and Categorization in Modular Neural Networks bills itself as being of interest to psychologists, cognitive scientists, neuroscientists, and to those interested in neural networks in general. For the most part, that assessment is right on the mark. Murre introduces and discusses in detail a biologically and psychologically plausible neural network. This is no small undertaking, and his results are impressive. However, despite the advertisement, his model has some distance to go before one could rightly claim that it is truly connected to either biology or psychology. Here I list a few of these shortcomings (some of them should be quite familiar); I then close by indicating what -- despite these shortcomings -- I find quite interesting and exciting about the CALM ("categorization and learning in neural networks") family.
2.1 Of course, Murre does not claim to have captured all the aspects of cortex involving information processing. Like most modelers interested in biological realism, he focusses on a few salient facts about our brain. In this case, he highlights the modularity of the minicolumns in neocortex and the organization of excitatory and inhibitory connections. He builds these two constraints into a network and notices that such a biologically inspired network can learn sequences of patterns unsupervised. There the biological connection ends, however. In fact, I find the actual connections among the various nodes in each module -- especially among the connections from the E-nodes to the R-nodes and from the V- and R-nodes to the the A-nodes -- to be only faintly inspired by any neuroscientific data, if at all.
2.2 This lack of biological realism should be easily forgiven once it is realized that Murre has no intention of modeling known neurophysiological phenomena. Instead, he wants to use his network to explore high level psychological curiosities, such as letter recognition, the difference between implicit and explicit processing, and pattern recognition in general. Thus, we are left to wonder why Murre would bother connecting his model to the few biological facts that he does, given that (surely) the minicolumn and local excitatory and inhibitory connections operate at a level of organization much lower than the information processing phenomena he discusses. The price of the chasm between the level of organization of the neurophysiological details he uses and the level of organization of the psychological phenomena is that his biological inspiration is no longer interesting as a design feature. Let me emphasize, though, that I do not wish to ignore the power of the model Murre has developed, nor do I wish to slight the implications it might have for rethinking psychological theories. What I do wish to claim is that the power and the implications have nothing to do with biology -- nor should they.
3.1 On the other hand, I take the claim for psychological plausibility to be more serious and far-reaching. Here, Murre's errors are more subtle, yet no less significant. Allow me to take his model of implicit and explicit processing as an example. He designed ELAN such that it would receive three types of inputs -- a context, a word beginning, and a word ending-- and then give one output -- a complete word. The task for the net after training is to output the correct completed word given either a word beginning and a context (explicit memory test) or just a word beginning (implicit memory test). Already the connection to language processing is dubious. For example, it is fairly well accepted that we process and store words in meaningful chunks that do not necessarily correspond to syllables or to word beginnings/word endings. The task of learning a two-part pattern in different contexts and then completing the pattern given the first half is a substantially different paradigm from the word completion task used in explicit and implicit memory tests, which traditionally ignores subitized parts when dividing letter strings.
3.2 Yet Murre seems to have overlooked something more damaging. He is interested in the differential processing between high frequency and low frequency WORDS. The data he reports, which reflect mainstream psychological opinion as far as I know, indicate that subjects who have been previously exposed to a list of both high and low frequency words show a greater effect in a word completion test for implicit memory for the low frequency words. However (and this is crucial), they still do show significant priming effects for the high frequency words. Murre reports that "[t]he average increase in completion performance compared with base rate for the low-frequency words... was significantly greater than that for the high-frequency words" (p. 78). What he fails to report, however (perhaps because he did not notice), was that his neural net "subjects" did not show ANY significant priming effect for the high frequency words. (The average base rate of completion for high frequency words was 5.67 and the experimental word completion performance was 6.75. According to a quick t-test, this difference is insignificant (t(11) = -2.6).) His network fails to model the psychological phenomena accurately.
3.3 What this failure tells us is that the word completion task is too easy for Murre's "subjects" in that they do not need to rely on context to determine the correct answer. Perhaps this is an artifact of the artificial word-beginning/word-ending division he has imposed. Perhaps if he were to change the way the inputs were given to the nets so that they are trained on complete words and a context and then tested on partial patterns (with and without a context input), the results would be more accurate. At least the inputs would then be psychologically more plausible.
4.1 Despite the overstated claims of biological and psychological plausibility, however, Murre's proposed models are nonetheless impressive. For example (and bugs aside), he does illustrate a way in which our memory systems might be constructed such that implicit and explicit memory processes might operate over the SAME mnemonic system. These sorts of existence proofs for various psychological or biological hypotheses I take to be the greatest contribution that neural network modeling can make for those interested in how we actually cogitate.
4.2 Moreover, Murre's CALM family of models manages to learn these sorts of patterns without direct supervision. Removing the dependence of neural nets on such highly artificial and obviously implausible learning algorithms as back-propagation and yet still managing to obtain sophisticated results is a feat for which Murre should be applauded. Though the ties to biology may be questionable, he has at least hit upon a modeling scheme that allows for true biological connections to be forged.
4.3 In sum, I recommend this book for those who wish to be up-to-date on the latest in connectionist technique. Murre gives much detail about the construction of the CALM modules, convincingly demonstrates their power through a wide range of applications, and gestures towards more impressive results yet to come. (However, for those new to the field, reading an introductory text on exactly what parallel distributed processing is and how it works is advised before plunging into the details of CALM.) But, if you are looking for a model of how cortex might actually work or of how we might actually think, then your expectations are too high. Murre's CALM is but a step in either of those directions.
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