Signal Processing via Neural Networks become practical -
via Analogic TeraOPS Visual Microprocessor Chips
Monday December 11, 2000, Morning
Recently, Intel shipped the first Tera FLOPS supercomputer consisting of almost ten thousand Pentium microprocessors. In many image processing applications we really need this trillion operations per second, however, the operations are special and do not require the 32 bit floating point accuracy. One alternative is the analogic CNN array computer performing about Tera equivalent operations per second, however, on a single chip. Indeed, the very recent stored programmable 64x64 processor CNN Universal Machine Chip [4] made in Seville with focal plane optical input, has a 3TOPS equivalent on chip computing power.
Behind this enormous signal processing computing power, there is a revolutionary computing paradigm: the Analogic Cellular Computer architecture [3] combining analog array dynamics with logic, all in a stored programmable framework.
What has been missing in many excellent neural network architecture? It seems, that stored programmability has been the real missing element, which has been the key for success in case of the digital microprocessors.
Ten years ago, in the seminal, paradigm forming, and now historic paper L.O. Chua and L. Yang [1] introduced the Cellular Neural/nonlinear Network (CNN). It is a 2D or 3D regular array of locally interconnected nonlinear dynamic systems called neurons, or cells, whose global functionality is determined by a small number of parameters. These parameters define the local interconnection pattern, called cloning template. Once the cell is given, the cloning template, specify the operation of the whole array. The cloning template is the protagonist in CNN. It is like a gene for spatio-temporal dynamics. Using very simple cells, even first order ones, practically all the simple and exotic spatio-temporal nonlinear dynamic phenomena can be generated by “engineering” the cloning template. Like genes, cloning templates can define a whole universe of phenomena. Designing this template we can engineer this universe. It was shown quite early that many neuromorphic models of the visual pathway can be represented by CNN models [5]. Due to sparse, mainly local connectivity, CNN is very convenient in VLSI design.
The invention of the CNN Universal Machine architecture [3] put the CNN dynamics into a different perspective. The CNN spatio-temporal dynamics, via the cloning template, became the atom, the elementary instruction of a stored program in this new computational paradigm of Spatio-temporal Instruction Set Computers (StISC). A new world of analogic algorithms and software has been developing.
In this lecture, the architecture, the physical implementation, the application, and the biological relevance of this new technology will be presented.
References