Abstract
Many real-world mission-critical applications require continual online learning from noisy data and real-time decision making with a defined confidence level. Brain-inspired probabilistic models of neural network can explicitly handle the uncertainty in data and allow adaptive learning on the fly. However, their implementation in a compact, low-power hardware remains a challenge. In this work, we introduce a novel hardware fabric that can implement a new class of stochastic neural network called Neural Sampling Machine (NSM) by exploiting the stochasticity in the synaptic connections for approximate Bayesian inference. We experimentally demonstrate an in silico hybrid stochastic synapse by pairing a ferroelectric field-effect transistor (FeFET)-based analog weight cell with a two-terminal stochastic selector element. We show that the stochastic switching characteristic of the selector between the insulator and the metallic states resembles the multiplicative synaptic noise of the NSM. We perform network-level simulations to highlight the salient features offered by the stochastic NSM such as performing autonomous weight normalization for continual online learning and Bayesian inferencing. We show that the stochastic NSM can not only perform highly accurate image classification with 98.25% accuracy on standard MNIST dataset, but also estimate the uncertainty in prediction when the digits of the MNIST dataset are rotated. Building such a probabilistic hardware platform that can support neuroscience inspired models can enhance the learning and inference capability of the current artificial intelligence (AI).
Introduction
Modern artificial intelligence excels at pattern recognition but often struggles with two critical requirements: continual learning—the ability to learn new information without catastrophic forgetting—and uncertainty estimation. In biological systems, synapses are inherently stochastic, with signal transmission probabilities often ranging between 10% and 50%. While traditional engineering views such variability as a nuisance, neuroscience suggests this noise is a functional feature that helps the brain escape local minima and perform probabilistic reasoning. Replicating this "always-on" noise in conventional silicon architectures is computationally and energy-intensive. This work presents a hardware-driven approach to probabilistic computing by building a "Neural Sampling Machine" (NSM) that mimics biological synaptic stochasticity directly within the hardware fabric.
Methods and Hardware Implementation
The researchers developed a hybrid stochastic synapse integrated into a crossbar array, consisting of two primary components: a Ferroelectric Field-Effect Transistor (FeFET) acting as analog memory, and a two-terminal stochastic selector (Ag/HfO2) acting as a noise generator.
By pairing these components in series, the synapse performs Bernoulli sampling of the weight. This hardware-level implementation of "DropConnect" allows the network to sample from a distribution of weights naturally. The system was evaluated using the standard MNIST dataset for classification benchmarking and a rotated MNIST dataset to test uncertainty quantification.
Key Insights: Noise as a Regularizer
A central finding of this study is that the multiplicative noise injected by the hardware selector acts as a continuous mask, forcing the network to learn robust features. This stochasticity leads to autonomous weight normalization, a process that typically requires expensive computation (e.g., Batch Normalization) in software.
Furthermore, because the stochasticity is "always-on," the NSM performs Monte-Carlo sampling during inference. Instead of a single deterministic output, the system produces a probability distribution, allowing it to quantify its own confidence.
Results
The simulated NSM achieved a classification accuracy of 98.25% on the MNIST dataset, matching the performance of state-of-the-art deterministic models. More importantly, the NSM demonstrated superior uncertainty quantification. When presented with rotated digits (out-of-distribution data), the system showed a significant increase in predictive entropy, effectively signaling its own confusion. Additionally, the NSM maintained a stable weight distribution throughout the learning process, confirming the effectiveness of the hardware-driven weight normalization.
Discussion and Future Directions
While the NSM offers a promising path toward brain-like AI, challenges remain regarding device endurance. The current Ag/HfO2 selectors are rated for approximately $10^8$ cycles; intensive on-chip training may require materials with higher endurance, such as Vanadium Dioxide. Future work will also explore local learning rules to replace global backpropagation and the development of fully integrated physical crossbar chips to realize the projected energy efficiency gains.
Summary updated