A high-speed analog neural processor

Abstract

Targeted at high-energy physics research applications, our special-purpose analog neural processor can classify up t o 70 dimensional vectors within 50 nanoseconds. The decisionmaking process of the implemented feedforward neural network enables this type of computation to tolerate weight discretization, synapse nonlinearity, noise, and other nonideal effects. Although our prototype does not take advantage of advanced CMOS technology, and was fabricated using a 2.5-pm CMOS process, it performs 6 billion multiplications per second, with only 2W dissipation, and has as high as 1.5 Gbyte/s equivalent bandwidth.lthough neural networks offer exceptionally powerful parallel computation performance, most current applications focus on exploiting their leaming capabilities. The ability of neurai networks to learn from examples has given rise to several quite successful experiments. Those involving handwritten character recognition, speech recognition, and similar challenges in which biological systems overshadow artificial intelligence come to mind.Still, the benefits of unique parallel processing that neural networks afford warrant our attention as well. With fully parallel neural hardware, processing time is independent of the data-size the network must process. Only a few computing steps require serial processing, making computation time extremely short. The inner product computation involved does present one major challenge for realizing the hardware of neural nets. Therefore, if an application does not demand high precision, the compact, high-speed analog approach provides great advantages.Analog techniques let us create single-chip architectures of complex neural networks, featuring low cost and low power dissipation. Such hardware, offering processing times as low as several microseconds for as large as 128 dimensional input vectors, is already commercially available.' Still, solving for application domains that demand tens of nanoseconds of processing delayzt3 for similarly large input vectors is almost impossible. The new architecture this article describes provides precisely this sort of highcomputing performance, offering one solution for a number of demanding applications. Nuclear research applicationsTo help understand the behavior of fundamental particles and forces, Hamburg's High Energy Physics (HEP) Institute Deutsches Elektronen Synchrotron (DESY) operates two large detectors installed within its hadron-electron ring accelerator (HEM). They are called H1 and Zeus. The two detectors contain different components, each specialized for detecting track, momentum, or energy of particles coming from the interaction region, where electrons and protons collide.These detectors provide tremendous amounts of information through 200,000 analog channels, sampled at a rate of 10 million times a second and producing 10l6 bytes of data per second. The resulting data flow, which requires real-time processing, imposes a great challenge for the dataacquisition system, far exceeding the capabilities even of available superco...

Winner-Take-All Network Utilising Pseudoinverse Reconstruction Subnets Demonstrates Robustness on the Handprinted Character Recognition Problem

“…Moreover, this property allows for on-chip learning. Consider, for example, mixed (analogue and digital) VLSI chips [18][19][20] that utilise bit valued memories for each connection in an analogue network, and can set the bit values from outside. The dynamic pseudoinverse procedure suggests that a similar construct might suit the requirements of the pseudoinverse computation with highly relaxed requirements on connection strengths.…”

“…Here, a categorising architecture is suggested that, accordingly, exhibits 'categorical perception' sometimes at the expense of increased reconstruction error. Also, the training procedure utilised here is adjusted to the needs of mixed VLSI technologies exhibiting on-chip learning [18][19][20]. The training method features the attractive properties of ART, i.e.…”

A Survey of Neuromorphic Computing and Neural Networks in Hardware

“…In fact, the original neuromorphic definition by Carver Mead referred to analog circuits that operated in subthreshold mode [1]. There are a large variety of other neuromorphic analog implementations [7], [8], [12], [14], [17]- [19], [99], [100], [323], [324], [326]- [329], [331]- [333], [335]- [372], [374]- [387], [516], [571], [572], [578]- [580], [594]- [603], [606]- [610], [612]- [614], [620]- [640], [642]- [646], [648]- [653], [655]- [664], [666]- [683], [685]- [699], [701]- [703], [705]- [719], [722]- …”

Neural Networks in Analog Hardware — Design and Implementation Issues

“…A chip targeted at high-energy physics research applications was proposed by Masa 130 and is presently available as a commercial product. The chip can classify vectors of up to 70 components within 50 ns.…”