Browse Books

Reviewer: Jose Hernandez-Orallo

While the title may lead to thinking that this is a book about the connection between neural networks and intellect, Perlovsky presents instead a very personal view about philosophy of mind and the mathematics behind Modeling Field Theory (MFT), the authors theory of learning based on adaptive fuzzy models given a priori, where parameters and fuzzy membership are learned iteratively. It is important then to remark that the book just deals with a particular kind of neural networks, a priori model-based neural networks. Other kinds of neural network algorithms are neglected or just assumed as known. The book combines quite philosophical stuff with complex mathematical concepts. While some parts are suitable for a broad audience, many other parts are really esoteric, and few readers will get the most of the book in a first reading. The first part is a lengthy introduction and motivation for the rest of the book. Some of his theses are introduced, such as that (some) knowledge has to be given to us a priori, the mathematics of learning instinct is related to the concept of beauty, the need of fuzzy logic to avoid computational intractability, and other more misleading assertions such as an emergence of ordered cosmos is equated with the divine act of creation, which psychologically is equivalent to an emergence of consciousness. The so-called 2300-year gap between Plato and Aristotle and modern theories of mind is exaggerated, and rediscovers the modern discussion about the proportion and relevance of a priori (innate) wrt. a posteriori (learned) knowledge. In this regard the author does not explain why humans have less predetermined (innate) and more adaptive behaviors than other animals. His (unsupported) explanation is that humans have a greater number of a priori (and more general) models. His thesis that the nominalist philosophy...cannot explain learning of complex concepts by mind ignores the possibilities of incremental learning. The second part is quite technical. The presentation is formally irreproachable, but it lacks the didactical focus of a textbook. For instance, chapter 4, which describes the major issues of MFT, could include an example about the use of parameters. Moreover, although this chapter is essential, it is not included in any of the six proposed course outlines in the preface, maybe suggesting its difficulty. Formal proofs of convergence are provided but the supposed efficiency of MFT is only shown through examples. While chapter 5 is about mixture models (Maximum Likelihood Adaptive Neural System, MLANS), where parameters describe means and variances, chapter 6 works with mixture spectrum models (Einsteinian Neural Networks, ENN), where parameters may be spectral values, e.g. frequencies. Although MLANS and ENN are compared with nearest neighbor and statistical methods, a broader comparison with other more recent ML methods would make statements such as MLANS is an efficiently network and ENN learning is fast and efficient more credible. Chapter 7 presents more specific applications of the previous methods. Chapter 8 extends MFT for quantum computation. In chapter 9, Cramer-Rao bounds are shown. More recently shown limits of learning are not discussed. Finally, a computer scientist may find chapter 10 nonsense, especially his explanation of beauty, as the author almost agrees a reader might exclaim: you dragged me through four paragraphs of high-flying notions.... The third part is intended for fun and is highly lucubrative. Chapter 11 discusses (and discredits) recurrent opinions about the impossibility of machine intelligence: supposed implications of Godel theorems and famous Penroses ideas. Chapter 12 includes a rather classical review of the conception of consciousness, except from the references to Buddhism, the author’s cat and his conclusion: I consider the only hard questions about consciousness to be free will and the nature of creativity. The discussion about collective and individual consciousness of some cultures is ungrounded, and scientifically unacceptable attitudes are shown: if science cannot explain free will, most will doubt this aspect of scientific thought, rather than free will. Therefore, science either has to explain free will, or to acknowledge that here is the boundary of applicability of the rational method as far as it exists today. From a more general standpoint, the book could certainly be enriched with Solomonoff’s theory of learning and intelligence, the modern view of information based on Kolmogorov complexity and Chaitin’s ideas about the limits of mathematics and Godel’s incompleteness theorems. A persistent issue the reader may want to be answered is how these a priori models appear. The treatment of this topic is exceptionally superficial, and is left with open questions. Obviously, a book of this kind has to end with more questions than answers, but the answers are mostly based on intuition. This is one of the main weaknesses of the exposition: the author does not clearly draw the line between scientific results and opinions, even though he says he does. Online Computing Reviews Service