This volume grew out of seminal papers published in IEEE Transactions on Circuits and Systems [1,2], as well as from lecture notes from courses conducted at both Berkeley and Budapest over the past six years. Analogic cellular neural networks (CNNs) are fully connected arrays that mimic the physiology of human sensory and processing organs, and incorporate local storage. The authors tout CNNs as “the first operational, fully programmable industrial-size, brain-like stored-program dynamic array computer,” able to perform the equivalent of three trillion digital operations per second (and thus outperforming digital signal processors (DSPs) by over three orders of magnitude, as measured in terms of speed, power, or area).
Chapter 2 introduces the basic architecture, notations, definitions, and mathematical foundations of CNNs, and highlights the importance of the local interconnection synaptic weight (cloning template or gene). These templates define the array dynamics that generate an output image from an input image. TEMLIB, a CNN template library, is included in Appendix A.
Chapter 3 introduces a simple technique for determining these array dynamics, based on cell dynamics, and describes 11 useful templates.
The next chapter covers digital computer simulation of CNN dynamics (with the accompanying CANDY simulator included in Appendix C). Numerical integration algorithms, digital hardware accelerators, and analog implementations are also discussed.
In chapter 5, the characterization of the simplest form of CNN is explored, and the binary input/binary output case described. We learn that a basic 3 × 3 template is capable of performing 2512 different local Boolean functions!
A unified theory of uncoupled CNN templates appears in the next chapter, while the CNN universal machine architecture is introduced in chapter 7. Each cell in the latter comprises not only a local logic unit, but also local analog and logic memory, together with global clock, thus enabling the realization of every local Boolean function.
A mathematical analysis of CNN stability, in terms of cloning templates, is presented in chapter 8. Chapter 9 expands into a complete CNN architecture, and accompanying high level language, compiler, operating system, and development environment.
Chapter 10 covers template design and optimization algorithms; and the accompanying TEMPO program is included in Appendix B. Two-dimensional CNN linear filters and the discrete space Fourier transform are discussed in chapter 11. In chapter 12, a simple coupled CNN is applied to the global connectivity problem (one which has long been considered impossible on locally connected arrays).
Nonlinear and delay type synaptic weights and their uses are covered in chapters 13 and 14, respectively. Complementary metal-oxide semiconductor (CMOS) analog and digital implementations of the CNN universal machine are discussed in chapter 15. The final chapter highlights the similarity between CNN architectures and models of the visual pathway, and compares models and measurements of living retina.
Chua and Roska make the observation that “in modern image processing, PDE-based techniques are becoming the most challenging and important new directions.” They point out that these are the native, elementary instructions of analogic CNN computers (just as add, multiply, nand, and xor are of digital computers), and welcome the challenge of devising useful algorithms to run on their standard spatio-temporal CNN universal machine-based computer. Their goal is also to build analogic adaptive sensor-computers, in which sensing and computing are fully and functionally integrated on a single chip. In true evangelical spirit, they claim that “a new understanding about computing itself is emerging!” The jury is still out, but I cannot help but wish them well in their quest, given the innovative nature of their approach.
