Jeffrey G. Bernstein

Low power logic for statistical inference

Efficient hardware implementations of statistical inference continue to grow in importance for a wide range of computing applications. While CPU cycles are increasingly being used for statistical inference, transistors are also becoming increasingly statistical. For implementing statistical algorithms, could it be that statistical electronic substrates are a feature rather than a bug? We show that inference models can often be built from local constraints, and explain the gate-level mathematical functions required for the resulting inference solver. We suggest that signals should consist of probabilistic populations of particles representing samples from a probability distribution, with gate functions acting to transform these ensembles. Using this mapping from statistical physics to statistical inference, we present Bayesian logic circuits as highly efficient alternatives to digital standard cell libraries. For particular inference computations, novel VLSI architectures based on Bayesian logic circuits consume orders of magnitude less power and silicon area compared to conventional digital processors.

CONVOLUTIONAL NEURAL NETWORK

In Chapter 8, we looked at a traditional neural network (NN). One of the limitations of a traditional NN is that it is not translation invariant—that is, a cat image on the upper right-hand corner of an image would be treated differently from an image that has a cat in the center of the image. Convolutional neural networks (CNNs) are used to deal with such issues.