Haken’s monograph does not follow the mainstream of today’s neurocomputing literature. The title is somewhat misleading in that no exotic architectures are discussed. What is discussed can be implemented on any serial computer, or more efficiently on parallel computers, as is the case for traditional neurocomputing techniques. The difference comes from the author’s interpretation of pattern recognition as almost the same as pattern formation, and from his application of the power of synergetics to the recognition problem.
Readers unfamiliar with the concept of synergetics will learn that it is “an interdisciplinary field of research concerned with the spontaneous formation of spatial, temporal or functional structures by self-organisation.” Major issues of synergetics are discussed in the first four chapters, but this introduction is brief and more of a reminder than a full overview.
Chapter 5 gives the mathematical foundations of the formation of prototype and test feature vectors and other important clues in the process of recognition in a rigorous manner. Chapter 6 presents some recognition experiments. Network realization, the invariance of recognition results under different geometric transforms applied to visual patterns, and learning algorithms are dealt with in the next five chapters. One of the learning relations is recognized as a mathematical sharpening of the well-known Hebbian rule from traditional neurocomputing. As with all issues in this book, these five chapters are well illustrated with visual pattern recognition experiments and results. Next is a chapter that compares human and machine perception.
Chapter 13 treats the interesting problem of the perception of ambiguous patterns. This issue is important because real life is full of ambiguous situations. Chapter 14 is devoted to motion pattern recognition, an area that is in the beginning of its development. In chapter 15, the classical XOR problem is solved in the light of synergetics. One can easily see that the synergetic approach has no such things as spurious states, as in the now classical back-propagation, although the basic idea of attractor states used is about the same. Chapter 16 uses difficult mathematics, studying how mathematical models that reproduce cognitive abilities can be linked to the properties of biological neurons. Chapter 17 concludes the book and traces the way to some open questions such as stereo vision and motion perception.
This book is a mix of philosophy and mathematics. Surprisingly, it can be read without a strong mathematical background because key ideas are laid out clearly. Nonetheless, a reader whose intent is to experiment further based on the content of the book will be faced with high-level mathematics.