OscNet: Machine Learning on CMOS Oscillator Networks

TL;DR

OscNet introduces a brain-inspired CMOS oscillator network that minimizes a Potts Hamiltonian using forward propagation and Hebbian updates, offering a highly energy-efficient alternative to backpropagation-based ML. The framework supports MIMO configurations for image convolution, models prenatal visual-system development, and extends to unsupervised autoencoding, K-means clustering, and linear regression via forward-only learning. Key contributions include hardware-compatible oscillator dynamics, a phase-based data representation, and demonstrated MNIST performance that surpasses certain autoencoder baselines while enabling compact, energy-efficient training. The work suggests OscNet could serve as a practical computational fabric for next-generation AI systems, balancing biological plausibility with competitive task performance.

Abstract

Machine learning and AI have achieved remarkable advancements but at the cost of significant computational resources and energy consumption. This has created an urgent need for a novel, energy-efficient computational fabric to replace the current computing pipeline. Recently, a promising approach has emerged by mimicking spiking neurons in the brain and leveraging oscillators on CMOS for direct computation. In this context, we propose a new and energy efficient machine learning framework implemented on CMOS Oscillator Networks (OscNet). We model the developmental processes of the prenatal brain's visual system using OscNet, updating weights based on the biologically inspired Hebbian rule. This same pipeline is then directly applied to standard machine learning tasks. OscNet is a specially designed hardware and is inherently energy-efficient. Its reliance on forward propagation alone for training further enhances its energy efficiency while maintaining biological plausibility. Simulation validates our designs of OscNet architectures. Experimental results demonstrate that Hebbian learning pipeline on OscNet achieves performance comparable to or even surpassing traditional machine learning algorithms, highlighting its potential as a energy efficient and effective computational paradigm.

Figures (4)

• Figure 1: The Oscillator Network (OscNet) is a CMOS circuit inspired by the brain, where current serves as the medium for information transmission between oscillators. The network employs the Hebbian rule for learning, enabling adaptive and efficient connectivity among oscillators. Like the brain, OscNet calculates model responses and update weights with forward propagation only. Compared to recent deep learning models where back propagation is also needed, OscNet saves time and energy.
• Figure 2: Modeling of human visual perception. Light from the real world travels to the retina, where it is processed by retina cells. These cells are densely connected to the LGN. The retina generates waves that help the brain optimize the weights $w_{ij}$ so that $l_j = p_j$, even without explicitly knowing the positions of the retina cells or having seen any real-world images beforehand.
• Figure 3: MIMO OscNet simulation for 4 inputs and 3 outputs.
• Figure 4: Learned features of OscNet and Auto Encoder on MNIST with 10 hidden neurons.