Learning and decision‑making in animals are often modeled as Bayesian processes, where sensory evidence is combined with prior beliefs to guide behavior under uncertainty. The essential question is how noisy neural and synaptic dynamics instantiate this capability. This article presents a biologically grounded framework in which stochastic sampling from an internal energy function enables inference and learning. Variability of neural activity encodes uncertainty over latent states, while synaptic fluctuations represent posterior distributions over model parameters.
Within this framework, predictive coding networks can express epistemic uncertainty through Markov chain Monte Carlo (MCMC) sampling. The sampling mechanism mirrors intrinsic noise in biological systems and parallels electrical noise in emerging probabilistic analogue memory technologies.
Analog in‑memory computing hardware, built on probabilistic analogue memory devices, can perform massive, energy‑efficient probabilistic inference directly in hardware: the noise inherent to the devices serves as the source of samples, eliminating the need for separate random‑number generators. Consequently, analogue in‑memory computing emerges as a natural solution for scalable Bayesian inference.
Review