Emerging “sensory swarm” and “edge computing” paradigms rely on local information processing by resource-constrained nodes, thus reducing the required communication with central cloud infrastructure. Stochastic Computing (SC) offers extremely compact, error-tolerant and power efficient implementations of certain complex functions, at the expense of longer computation times and some degree of inaccuracy. Initially introduced as a technology “in between” analog and digital circuits, SC has experienced a revival in the last few years. Breakthroughs are being reported in applications such as digital filters, image-processing algorithms, Bayesian inference, LDPC decoding, and neural networks.
The presentation will demonstrate combinational and sequential SC primitives, their integration into a larger system, and interfacing between stochastic and binary domains. We will discuss SC’s advantages, drawbacks and potential solutions to overcome these drawbacks. Finally, we will give a brief overview of somewhat more futuristic SC applications: direct processing of neural signals; straightforward integration with memristive technology; and neural networks secured against adversarial attacks.