The cognitive brain

The cognitive brain

Book reviews 225 inside. Very good general presentations and valuable data provided in some chapters neighbour trivial examples and information of p...

239KB Sizes 0 Downloads 45 Views

Book reviews

225

inside. Very good general presentations and valuable data provided in some chapters neighbour trivial examples and information of problematic applicability. The book comes with a floppy disk with simple programs (no GUI, no mouse support, no hot-keys, DOS version only) and contrary to the cover description it is not user-friendly (no help function). There are also bugs in the software, for example after installation a message ‘invalid directory’ comes up and sometimes (caused by bad input) the computer hangs up. Most of this criticism came from my students who were asked to use it in a lab. Of course, one can still use the software to perform simple experiments with CNNs but not for an in-depth learning and certainly not for any practical application. In conclusion, the book cannot be used as a tutorial on cellular neural networks because of a lack of systematic presentation of the basic ideas. Also, the same information is repeated in many chapters written by different authors. An example is the description of the cellular neural networks principles, network organization and templates. The editors of the book cannot be blamed for all its faults since it is an edited version of the special issue of the International Journal of Circuits Theory and Applications containing papers chosen from the First IEEE International Workshop on Cellular Neural Networks and their Applications held in Budapest in 1990. Certainly, the problems of Cellular Neural Networks design, programming, and applications are important, interesting, and actual. But some of the papers which were of interest in 1990 may no longer be so in 1995. This book will not be very useful for scientists and engineers who want to apply neural networks in practice. For students, the book may be difficult to use because of the lack of coherent presentation of the material. Ryszard Tadeusiewicz Director, Biocybemetics Laboratory, Institute of Automatics, AGH Technical University, Krakow, Poland

The Cognitive Brain, by Arnold Trehub. MIT Press, 1991, ISBN O-262-20085-6. The book consists of an approach to modelling and implementing a series of cognitive functions through artificial neural networks. In addition to its unusually comprehensive scope, Trehub’s book distinguishes itself from many of the books in artificial neural networks that have been flooding the market by its consideration and adoption of modern results from neurophysiology, neuroanatomy and psychology. Selective attention, autaptic neurons, geometric transformation invariance, modular and topographical organization are some of the many modern and important paradigms that have been addressed. The Cognitive Brain is composed of sixteen chapters plus bibliography, which are briefly reviewed in the following. Chapter 1, Introduction, presents the basic underlying features adopted throughout the book. It includes an interesting position according to which the three levels of explanation characterizing artificial information-processing should be understood jointly, and not separately as originally proposed by D. Marr [2]. By considering the properties of the biological neural hardware, the possible high-level

226

Book reviews

models can be constrained to their most feasible possibilities. Other interesting issues aptly treated in this chapter include the identification of the characteristics that should be expected of a sound cognitive model, the justifications for adopting the comb filtering paradigm in detriment of Parallel Distributed Processing as well as discussions about representation, learning and the identification of the tasks to be performed. The second chapter specifies the assumed basic biophysical properties of neurons, discusses how long-and short-term learning and memory can be achieved. The adopted long-term potentiation mechanism involves the density of the axon transfer factor (ATF) and dendrite transfer factor (DTF), which implements a kind of Hebbian learning in which the synapses that present correlated activity are strengthened. Short-term memory is implemented by using autaptic cells, a special kind of neuron that is characterized by positive feedback through recurrent collaterals of its own axon. The objective of Chapter 3 is to introduce one of the major building blocks to be adopted in the neural models developed along the book. It consists of the synaptic matrix, composed of the imaging matrix and the detection matrix. Input is fed into the imaging matrix and the output is derived from the detection matrix, whose final stages operate according to the winner-takes-all fashion. By feedbacking the output lines into the imaging matrix, associative recall can be achieved. Synaptic matrices, the adopted basic neural mechanisms for learning, memory, and imagery, are able not only of learning and recalling static representations, but also of learning sequences of events. Synaptic matrices can be combined as a means of achieving more sophisticated behaviour. The second major neural structure adopted in Trehub’s framework consists of the retinoid (Chapter 4), a dynamic buffer which includes a series of retinotopically structures autaptic cells. Retinoids can be combined to yield retinoid systems, which typically receive and send topographically mapped projections. A retinoid is basically capable of parsing, constructing three-dimensional representations, locating and representing the reference (self) respectively to a specific scene, and performing selective attention, which is achieved through the special self-locus retinoid. Retinoid variants are developed which are capable of representations invariant to geometrical transformations and stereo processing. Some complementary underlying issues such as afferent-field aperture, novelty detection, scale and rotational invariance, clocking and episodic processing are discussed in Chapter 5. The synaptic matrix and the retinoid are then used by Trehub as a means of implementing a large variety of cognitive functions, including semantic networks (Chapter 6), analysis and representation of object relations (Chapter 71, composing behaviour (Chapter 81, formation and resolution of goals (Chapter 9), learning and recognition of simple visual patterns (Chapter 101, recognition of objects with varying shapes (Chapter 111, learning in complex environments (Chapter 121, narrative comprehension (Chapter 13) and illusions (Chapter 14). Chapter 15 provides an interesting substantiation of the adopted models and paradigms from the perspective of recent clinical findings. The last Chapter provides an overview of the reported contributions.

Book reviews

227

All in all, The Cognitive Brain reports an interesting approach to the development of cognitive models in terms of artificial neural networks. Unlike so many alternative approaches to neural models, Trehub’s book includes a series of important findings from neuroscience research. One of the few drawbacks of the book consists of the lack of a formal mathematical assessment of the neural models adopted. Also, the local code (comb filtering) paradigm has been recently criticized in favour of distributed coding (see e.g. [3]). The Cognitive Brain also presents some interesting implications from the perspective of consciousness research, as has been discussed in [4]. References [l] A. Trehub, The Cognitive Bruin (Cambridge, MA, MIT Press, 1991). 121 D. Marr, VZon (San Francisco, W.H. Freeman, 1982). 131 P.S. Churchland and T.J. Sejnowski, The Computational Bruin (Cambridge, MA, MIT Press, 1993). [4] L. da F. Costa, Getting the ghost out of the machine, Psyche (1039-723X/93) (1994).

Luciano da Fontoura Costa Cybernetic Vision Research Group IFSC - University of Sao Paul0 Caiva Postal 369 Sao Carlos, SP, 13560-970 Brazil Email: [email protected]. br

An Introduction to Biological and Artificial Neural Networks for Pattern Recognition, by Steven K. Rogers and Matthew Kabrisky, with contributing authors Dennis W. Ruck and Gregory L. Tarr. SPIE Optical Engineering Press, pp. 220, 1991, ISBN o-8194-0534-5, $30.00.

This book was published by SPIE in its tutorial texts series. Its text is intended for readers who wish to explore artificial neural networks (ANNs) for pattern analysis. The text starts with a lively presentation covering topics from the rich history of research of biological neural networks. The biological background emphasizes the connection between common artificial neural networks and biological neuronal processing. Visual information processing is emphasized with the goal of biological inspiration, not biological accuracy, of current artificial neural networks. Relevant biological parallels are presented throughout the text. An overview of artificial neural networks emphasizing the feedforward multilayer perceptron is provided next. Backpropagation is derived to allow the reader to vary any assumption about layers or neuron transfer function and develop their own learning algorithms.