Emergent Synaptic Plasticity from Tunable Dynamics of Probabilistic Bits

Sagnik Banerjee, Shiva T. Konakanchi, Supriyo Datta, Pramey Upadhyaya · 2025

Read Emergent Synaptic Plasticity from Tunable Dynamics of Probabilistic Bits by Sagnik Banerjee, Shiva T. Konakanchi, Supriyo Datta, Pramey Upadhyaya (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-00252.txt.

Introduction

Sagnik Banerjee and co-authors show in 2025 how noise can tune links in a probabilistic computer. Such machines use flickering bits that settle by chance, and they can crack hard problems, but setting the link strength between bits on a chip is awkward. The team hides an extra flickering bit inside each link.

By changing how fast the hidden bit flickers compared with the link's response, the effective link strength slides smoothly up or down. The tuning is directional: the pull from one bit to another can differ from the reverse pull. Math and simulations both confirm the effect. The authors say this learned link change, much like synapses in brains, could make such machines reprogrammable.

Catalog record

CREATORSagnik Banerjee, Shiva T. Konakanchi, Supriyo Datta, Pramey Upadhyaya
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-00252

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Emergent Synaptic Plasticity from Tunable Dynamics of Probabilistic Bits Preprint: APS/123-QED Sagnik Banerjee, Shiva T. Konakanchi, Supriyo Datta, Pramey Upadhyaya 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. I Introduction With the increasing data- and power-hungry computing demands in today’s world and the gradual decline of Moore’s Law Moore 1998 , domain-specific computing platforms have become popular Neuendorffer et al. 2021 . To that end, several approaches that utilize physics-inspired unconventional paths to computing have generated considerable excitement lately Finocchio et al. 2024 . This shift was envisioned by Richard Feynman in his seminal 1982 paper Feynman 1982 , where he proposed a quantum computer that can emulate and solve many interesting problems that involve complex probabilities and require path cancellation Shor 1997 . In the same paper, however, he also mentions a classical probabilistic (p-) computer that can deal with problems involving classical probabilities. Inspired by the same, a p-computer has recently gained attention Camsari et al. 2017 ; Datta 2024 ; Kaiser and Datta 2021 as a natural platform to efficiently emulate inherently stochastic problems. The fundamental building block of a p-computer is a stochastic unit called a p-bit, which produces a binary random output, m_{i} , determined probabilistically by an analog input, I_{i} . A p-computer is constructed by connecting the outputs and inputs of p-bits through synapses with coupling strengths J_{ij} . Such a p-computer naturally generates samples from a target probability distribution— encoded through the couplings J_{ij} — in 2^{n}- dimensional space with n p-bits. This feature lies at the core of efficient hardware emulation across a broad spectrum of problems—ranging from combinatorial optimization Aadit et al. 2022 and quantum emulation Camsari et al. 2019a ; Chowdhury et al. 2019 ; Chowdhury et al. 2023a on one end, to Bayesian inference Faria et al. 2021 and machine learning Niazi et al. 2024 on the other. The former benefits from interpreting m_{i} as an Ising spin. In this view, when J_{ij}=J_{ji} , the p-computer naturally generates samples from the Boltzmann distribution of an Ising spin system with energy E=\frac{1}{2}\sum_{ij}J_{ij}m_{i}m_{j} , coupled to a thermal bath Ackley et al. 1985 . In contrast, the latter leverages the interpretation of m_{i} as the output of a binary stochastic neuron, with the couplings J_{ij} functioning as synaptic weights Chowdhury et al. 2023b . Given the promise of p-computing, a wide range of physical platforms and phenomena—such as magnetic Borders et al. 2019 ; Hayakawa et al. 2021 ; Safranski et al. 2021 ; Shao et al. 2023 ; Rehm et al. 2024 ; Soumah et al. 2025 ; Lourembam et al. 2021 ; Duffee et al. 2024 ; Shukla et al. 2023 ; Misra et al. 2023 ; Rehm et al. 2023 , optical Roques-Carmes et al. 2023 ; Ma et al. 2025 ; Choi et al. 2024 ; Horodynski et al. 2025 ; Whitehead et al. 2023 , metal-insulator transition-based Seo et al. 2024 ; Bhaduri et al. 2025 , CMOS-based Li et al. 2025 ; Patel et al. 2024 ; Kim et al. 2024 and digital FPGA-based Pervaiz et al. 2019 ; Aadit et al. 2023 ; Aadit et al. 2022 ; Niazi et al. 2024 ; Nikhar et al. 2024 systems—have recently emerged as viable candidates for implementing p-computers. A functionality of critical importance for p-computers across these platforms is the ability to tune J_{ij} to generate desired probability distributions. So far, this has been largely achieved by integrating a network of p-bits with additional hardware Borders et al. 2019 ; Si et al. 2024 ; Mizrahi et al. 2016 ; Talatchian et al. 2021 ; Schnitzspan et al. 2023 ; Phan et al. 2023 ; Gibeault et al. 2024 specifically designed to tune J_{ij} . Simpler schemes allowing for on-chip tuning without the need for integration with additional components could offer more advantages. Engineering tunable couplings by modulating the intrinsic timescales of constituent elements have emerged as one such attractive mechanism across a broad range of unconventional computing approaches. For example, such approaches include tuning qubit frequencies to control interactions in quantum systems Yan et al. 2018 , modulating spiking rates in neuromorphic architectures Markram et al. 2012 , and adjusting oscillator frequencies in Ising spin implementations based on nonlinear oscillators Albertsson and Rusu 2023 . However, to the best of our knowledge, the question of whether tuning an intrinsic timescale of p-bits can enable programmable coupling between them remains largely unexplored. Figure 1: (a) Schematic representation of an interconnected p-bit network illustrating hidden p-bits (colored red) introduced between pairs of computational p-bits (colored black). A typical p-bit fluctuates randomly in time between +1 and -1. Each p-bit in the network receives continuous analog input I_{i} and produces a binary output m_{i} . The dynamics of each p-bit are characterized by two parameters: a correlation time \tau_{r}^{i} and a (k_{B}T)^{-1} like parameter \beta , that scales the non-linear activation function (\operatorname{tanh}) . The computational p-bits are characterized by correlation times \tau_{r}^{1} and \tau_{r}^{2} , while the hidden p-bit exhibits tunable stochasticity with adjustable correlation time \tau_{r}^{h} . The strength of synaptic coupling between m_{1{(h)}} and I_{h(2)} is labeled J_{1(2)} . All synapses are associated with a response delay \tau_{\mathrm{s}} . The effective coupling between m_{1} and m_{2} is depicted by J_{\mathrm{eff}} . (b) Programmability of J_{\mathrm{eff}} between the computational p-bits in an effective model by adjusting the correlation time of the hidden p-bit relative to the synaptic delay. In the hidden p-bit model, we set J_{1}\beta\ll 1 and J_{1}J_{2}\beta^{2}\gtrapprox 1 , thereby illustrating a directional asymmetry in J_{\mathrm{eff}} . This results in tunability of the interaction J_{21} (red) but not for J_{12} (blue). In this work, we propose and theoretically analyze a scheme to engineer effective couplings between p-bits by exploiting the tunability of their intrinsic fluctuation timescales. In natural systems influenced by random fluctuations, spatiotemporal noise injection is believed to interact with nonlinear dynamics of the system to give rise to counterintuitive regular behavior Sagués et al. 2007 . Motivated by recent developments in p-bit realizations driven by noise sources with controllable temporal correlations (see Konakanchi et al. 2025 ), we design a network in which the effective coupling between any two p-bits is tuned by inserting a third

Read the complete text