Stochastic Computing (scottlocklin.wordpress.com)

Stochastic computing — an idea dating to the 1960s that encodes numbers as streams of bits rather than precise voltages — is seeing renewed interest thanks to a recent startup push and refreshed access to older literature. Historically it competed with early supercomputers and analog machines for tasks like neural nets and control systems because streams of random (or merely uncorrelated) bits make arithmetic operations implementable with very simple, noise-tolerant logic. Early deployments included avionics, a small autonomous neural-robot, radar trackers and reservoir/spiking-net experiments; today the concept is being re-evaluated for modern ML workloads. The significance for AI/ML is practical and technical: stochastic representations match probabilistic and pulse-coded models naturally (spiking nets, graphical models) and allow trade-offs between accuracy, latency and energy by counting fewer bits or shorter streams. Core technical features include bitstream-based addition/multiplication via simple gates, memory implemented as delay lines, tolerance to noisy/transistor-level errors, and relaxed randomness requirements (uncorrelated streams suffice). Modern silicon — dense transistors, high-speed serial channels and voltage scaling — could make these architectures far more efficient than conventional digital accelerators; proponents claim orders-of-magnitude energy gains (aspirational today). If realized, stochastic hardware could offer a low-cost, inherently probabilistic substrate well suited to large-scale, noisy ML models and spiking architectures that are awkward to simulate on conventional processors.