Source-linked AI summary
Bio-inspired Learning and Decision-Making with Probabilistic In-Memory Computing Hardware: Part 1
Thomas Dalgaty, Eiji Kawasaki, Miguel de Prado, Devendra Vyas, Tommaso Salvatori
TL;DR
The paper asks how noisy neural dynamics can support Bayesian learning and decision-making and how those dynamics can be replicated in hardware. It proposes energy-based stochastic sampling over neural states and synaptic weights, then connects biological noise with analogue in-memory computing to support scalable probabilistic inference.
Problem
Predictive coding has underexplored epistemic uncertainty because many implementations treat synaptic weights deterministically or use variational posteriors.
Method
The framework uses stochastic neural and synaptic updates in an energy function, with Langevin dynamics sampling latent states and model weights.
Results
The paper presents biologically plausible MCMC inference and learning while identifying analogue in-memory hardware as a route to scalable, energy-efficient Bayesian computation.
Takeaways & Limitations
Intrinsic biological and device noise can be treated as a computational resource for probabilistic inference and uncertainty modeling.
Takeaways & Limitations
The predictive-coding neural update uses a transpose whose biological plausibility remains an open question.
Abstract
from arXiv · showhide
Learning and decision-making in animals are often modeled as Bayesian processes, where sensory evidence is integrated with prior beliefs to guide behavior in the face of uncertainty. But what are the inherent neural dynamics that give rise to this ability, and how could they be replicated in computing systems? This abstract discusses a biologically grounded framework in which noisy neural and synaptic dynamics perform inference and learning via stochastic sampling from an internal energy function, capturing uncertainty over latent states and model parameters through neural and synaptic variability, respectively. This enables approaches such as predictive coding networks to account for epistemic uncertainty via Markov chain Monte Carlo sampling. Drawing a parallel between intrinsic noise in biological systems and electrical noise in emerging probabilistic analogue memory technologies, we highlight how analogue in-memory computing hardware naturally emerges as the solution for massively scalable and energy-efficient probabilistic inference.
Introduction
The paper frames biological learning and decision-making as Bayesian inference implemented through energy-based, noisy neural dynamics. It identifies epistemic uncertainty in predictive coding as underexplored and proposes stochastic neural and synaptic updates as a route to MCMC-based inference.
- Animals are hypothesized to integrate sensory evidence with prior expectations through probabilistic reasoning under uncertainty.
- Energy-based models and predictive coding describe inference and learning through free-energy or prediction-error minimization.
- Predictive coding has limited treatment of epistemic uncertainty because many implementations use deterministic synaptic weights or variational posteriors.
- The proposed framework applies stochastic updates to neural and synaptic variables so inference and learning emerge as biologically plausible MCMC sampling.
Algorithm Formulation
The algorithm represents neuron states and synaptic weights probabilistically through an energy-based posterior and samples them using Langevin dynamics. It uses slower synaptic updates, parallel chains for uncertainty readout, and predictive-coding energy with an acknowledged biological-plausibility question.
- The posterior over latent neurons and synaptic weights combines likelihood and prior terms, whose logarithm corresponds to learning and latent-state energies.
- Langevin dynamics samples neuron-state and synaptic-weight distributions in a shallow linear predictive-coding network using prediction errors.
- Synaptic weights evolve more slowly than neural activity, and dataset learning scales the weight likelihood so samples reflect all observations.
- Parallel chains provide an instantaneous estimate of epistemic uncertainty through ensemble variability, avoiding the need to wait for one chain to mix.
- Predictive-coding energy is preferred to Hopfield energy because its dynamics are simpler and its shallow linear steady state is analytically obtainable.
- The neural update includes a transpose that analogue hardware can implement with forward and reverse crossbar reads, while biological plausibility remains open.
Probabilistic analogue in-memory computing
The paper motivates probabilistic analogue in-memory computing as a hardware route for accelerating intrinsically parallel MCMC inference. Device noise can be used to sample, store, and process probability distributions within memory rather than treating variability as a defect.
- MCMC with Langevin dynamics can require more than 200 iterations versus about 15 for a comparable deterministic predictive-coding solution.
- Analogue memory noise supports MCMC implementations that sample distributions, store samples, and perform vector arithmetic within the memory circuit.
- In-memory computation avoids the external DRAM bandwidth bottleneck that forces sequential parameter movement for massively parallel algorithms.
Research Outlook
The outlook is to scale energy-based predictive-coding models with probabilistic analogue in-memory hardware. It focuses on evaluating Bayesian predictive coding for learning and decisions and mapping its algorithm onto hardware's stochastic physics.
- The project aims to massively scale energy-based models such as predictive coding through probabilistic analogue in-memory computing.
- It will investigate Bayesian predictive coding for learning and decision-making within an energy-based framework.
- It will explore mapping the algorithm onto analogue hardware's intrinsic stochastic physics, reflecting variability observed in biological systems.