ZISC
Neural Networks on Silicon
Dateline: 07/18/00
Computing technology is a field full of surprises. Just as you thought that RISC (Reduced Instruction Set Computer) marked the road to follow, another abbreviation becomes a standard part of our everyday vocabulary. This time it is ZISC: Zero Instruction Set Computer. It is not a joke - it really does refer to a computer having "zero instructions", no matter how contradictory it may sound.
"Silicon" neural nets are old news, considering wide range of integrated neural networks based on analog technology. However, new, digital implementations allow for arbitrary precision, strong noise immunity and easy interfacing to other system components. ZISC technology was jointly developed by IBM and Guy Paillet, Chairman of Petaluma, CA-based Silicon Recognition, Inc. The first generation of ZISC chip contains 36 independent cells that can be thought of as neurons or parallel processors. Each of these can compare an input vector of up to 64 bytes with a similar vector stored in the cells memory: if the input vector matches the vector in the cells memory, the cell fires. Output signal contains the number of the cell that had a match or "no matches occurred" indicator. Simple enough, and pretty same as in the conventional serial machines. However, the parallelism is the first key to success - the speed of the system increases dramatically by eliminating the step of serial loading and comparing the pattern for each location. Another key factor is ZISC's scalability. A ZISC network can be expanded by adding more ZISC devices without suffering a decrease in recognition speed - networks with 10,000 or more cells are pretty common.
As for the low-level details, ZISC employs the Radial Basis Function (RBF) and K-Nearest Neighbor (KNN) algorithms. They determine how the network shall adjust the influence field of prototypes as new prototypes are stored in the network. It can be considered as an expert system that can recognize and classify objects or situations and take instantaneous decisions based upon accumulated knowledge. No "real" programming is required, as its generalization capability allows it to react correctly to objects or situations which were not part of the learning examples.
ZISC chips (PCI, ISA, PCMCIA and VME cards) are often used as a recognition engine plugged into a traditional computer. As they use very little power, ZISC cards can also be put into portable, stand-alone devices using 84-pin SIMs (Single Inline Modules) that contain either 3 or 6 ZISC chips with up to 216 processors. Todays ZISC chip contains more than 200 neurons per chip and can find a match among 1,000,000 patterns in one second operating at less than 50 MHz.
Practical uses of ZISC technology are focused on pattern recognition, information retrieval (data mining), security and similar tasks. Silicon Recognition was initially focused on few large customers with the objective to validate the concept of fully hardwired neural network chip. They are now moving full speed ahead to establish a worldwide presence through a network of distributors and partners. Developers of intelligent systems at ZoomSoft Corporation are reorganizing their focus with the advent of ZISC towards new interactive technologies for the next generation of computer systems and devices. According to the material available at the Web site, their are developing Companion OS, "the first sensory operating system of the post-PC era". It will use ZISC technology to perform tasks that are far beyond the reach of the current mainstream technology - face recognition at a rate of 15 frames per second without accessing the CPU, information retrieval out of terabytes of data within 3 micro-seconds, etc.
Pulnix America, Inc. recently announced ZiCam, a hardware neural network based zero instruction Smart Camera. It comes with a user-friendly Graphical User Interface called ZiCAM Teacher. Each inspection object is classified as a category in the neural network and is memorized. As part of smart camera training, the user relates the object with a predefined category. Custom modifications to the graphical interface can be easily implemented using Microsoft Visual Basic.
General Vision Inc. is also offering a wide range of ZISC-based devices and tools, including software development kits and evaluation systems. Their ZEFIR software for rapid prototyping of an object or pattern recognition application can be used in simulation mode for users without hardware neural network card. On a side note, Luca Marchese provides several software packages at ZISCTOOLS site. The NeuroSight I is a self-contained vision systems combining a camera and a neural network card for image learning and recognition, similar to the ZiCAM. Object inspection can be as simple as teaching the NeuroSight good and bad parts and letting it classify new parts per association - no programming necessary.
ZISC offers an extraordinary approach to the design of computing devices. Will it go beyond the existent bottleneck in standard computer architecture where the memory and processing unit are two different entities?