Mixed-signal VLSI implementation of the Products of Experts' contrastive divergence learning scheme

Fleury, Patrice; Murray, Alan F.

doi:10.1109/iscas.2003.1206397

Cited by 2 publications

(1 citation statement)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…There is still interest in developing neural network chips for algorithms: the Boltzmann machine's capability for generating representations (and for using noise in the algorithm itself) has led to continuing interest in that algorithm ( [77,78]. In addition there has been interest in hardware for the more recent products of experts algorithm [79], resulting in a mixed signal (hybrid) implementation [80].…”

Section: Developed Hardware For Time-free Neuronsmentioning

confidence: 99%

Implementing Neural Models in Silicon

Smith

Handbook of Nature-Inspired and Innovative Computing

View full text Add to dashboard Cite

Neural models are used in both computational neuroscience and in pattern recognition. The aim of the first is understanding of real neural systems, and of the second is gaining better, possibly brain-like performance for systems being built. In both cases, the highly parallel nature of the neural system contrasts with the sequential nature of computer systems, resulting in slow and complex simulation software. More direct implementation in hardware (whether digital or analogue) holds out the promise of faster emulation both because hardware implementation is inherently faster than software, and because the operation is much more parallel. There are costs to this: modifying the system (for example to test out variants of the system) is much harder when a full application specific integrated circuit has been built. Fast emulation can permit direct incorporation of a neural model into a system, permitting realtime input and output. Appropriate selection of implementation technology can help to make interfacing the system to external devices simpler. We review the technologies involved, and discuss some example systems. Why implement neural models in silicon?There are two primary reasons for implementing neural models: one is to attempt to gain better, and possibly brain-like performance for some system, and the other is to study how some particular neural model performs. Current computer systems do not approach brain-like system performance in many areas (sensing, motor control, and pattern recognition, for example, to say nothing of the higher level capabilities of mammalian brains). There has been considerable research into how the neural system attains its capabilities. Implementing neural systems in silicon can permit direct applications of this research by permitting neural models to run rapidly enough to be applied directly to data. It is true that increases in workstation performance have allowed some software implementations of neural models to run in real time, but the highly parallel nature of neural systems, coupled with increasing interest in the application of more sophisticated (and computationally more expensive) neural models has caused interest in more direct implementation to be maintained. Interest in applying neural models to sensory and sensory-motor systems has made attaining real-time performance a critical factor. Real sensory systems are highly parallel, with multiple parallel channels of information, so that even though each channel might be implementable in real-time in software, implementing multiple channels implies hardware implementation.The study of how particular models of neural systems perform is one aspect of computational neuroscience. Such studies are usually carried out in software, as this allows easy alteration of and experimentation with systems. However, models of the highly parallel architecture of neural systems run slowly on standard 1

show abstract

Section: Developed Hardware For Time-free Neuronsmentioning

confidence: 99%