Emergent Synaptic Plasticity from Tunable Dynamics of Probabilistic Bits

Abstract

Probabilistic (p-) computing, which leverages the stochasticity of its building blocks (p-bits) to solve a variety of computationally hard problems, has recently emerged as a promising physics-inspired hardware accelerator platform. A functionality of importance for p-computers is the ability to program-and reprogram-the interaction strength between arbitrary p-bits on-chip. In natural systems subject to random fluctuations, it is known that spatiotemporal noise can interact with the system's nonlinearities to render useful functionalities. Leveraging that principle, here we introduce a novel scheme for tunable coupling that inserts a ''hidden'' p-bit between each pair of computational p-bits. By modulating the fluctuation rate of the hidden p-bit relative to the synapse speed, we demonstrate both numerically and analytically that the effective interaction between the computational p-bits can be continuously tuned. Moreover, this tunability is directional, where the effective coupling from one computational p-bit to another can be made different from the reverse. This synaptic-plasticity mechanism could open new avenues for designing (re-)configurable p-computers and may inspire novel algorithms that leverage dynamic, hardware-level tuning of stochastic interactions.