As guest editors for the International Journal of Circuit Theory and Applications, we have the pleasure to introduce this second special issue on cellular neural networks. Cellular neural networks (CNNs) are mainly locally interconnected, regularly repeated, analogue (continuous-or discrete-time) circuits with a two-dimensional array or a three-dimensional lattice architecture. Embedded in the CNN universal machine architecture is the first stored programme continuous-valued array computer. Each cell or processor in a CNN is a non-linear dynamic system mainly coupled to its neighbours. Against a background of intensive worldwide research on artificial neural networks and parallel computer architectures, the CNN is now widely accepted as an ingenious and effective neural network and general dynamic 3D architecture. In fact, it makes a very interesting trade-off between four crucial issues: theoretical manageabilitiy , biologicial plausibility, ease of VLSI implementation and broad applicability. Briefly we believe that the CNN universal machine architecture is a prominent competitor in the ballpark between these four corner issues.
It was invented in a pair by papers of Chua and Yang in 1988'. Already three IEEE International Workshops on Cellular Neural Networks and Their Applications have been held (1990), 1992 Munich4. ', and 1994 Rome')). In addition, interesting special sessions or tutorials on CNNs have been held at two European Conferences on Circuit Theory and Design (1991, Copenhagen; 1993, Davos), two IEEE International Symposia on Circuits and Systems (1994, London, 1995, Seattle) and several other conferences.
On 18-21 December 1994 the Third IEEE International Workshop on Cellular Neural Networks and Their Applications was held in Rome. It attracted about 100 participants and the Proceedings' report the state of the research at the end of 1994 with contributions from 76 authors. Many fruitful discussions were held on these results and on open problems. Many new aspects and practical breakthroughs were presented, e.g. operational CNN universal chips in the framework of a CNN chip prototyping system, other innovative chips, learning procedures, global optimization, real life applications, the use of amorphous silicon technology, dynamic non-stationary operating modes in constructive use and the computational complexity of analogic CNN algorithms. In addition, it was shown how patterns of activity can be faithfully modelled physiologically and compared with measurements in living tissues. At the workshop it was considered to be very important to prepare an archival document on the state of the art in this area. Hence it was decided to launch a call for papers for this special issue, which was not only oriented to the authors of the workshop, but was also open to other researchers. Of the 40 received papers we expect to accept about 27, which will require two issues of the journal.
In these two special issues the papers are grouped around the same four topics: theory, biologically inspired architectures, VLSI implementations and applications. In the first issue we first find papers concerning important theoretical aspects such as stability criteria, design and learning, spatiotemporal phenomena, oscillations, wave propagation and spatial logic. On the topic of biologically inspired architectures, one paper deals with modeling aspects of stereo vision and its extension to 3D depth detection. In the part on VLSI realizations, two VLSI chip designs-an optical and binary input CNN universal chip and a controllable analogue input/output chip-are reported along with an interesting optoelectronic architecture. The last part deals with applications. Challenging image recognition tasks can be handled by CNN algorithms, thereby realizing a kind of 'bionic eyeglass'. Finally, CNN templates for