Integration of oscillator-based feature extraction for energy-efficient convolutional neural networks

The rapid growth in machine learning and artificial intelligence workloads has increased the demand for computing power. Although digital computing accelerators dominate today’s market, analog computing architectures are emerging as promising energy-efficient alternatives. In this work, we propose an approach to address these challenges by replacing the first layer of a convolutional neural network (CNN) with a network of oscillatory retinal neurons (ORNs) composed of weakly coupled, optically activated negative differential resistance (NDR) devices. Each ORN consists of a photodetector exhibiting NDR behavior under illumination, coupled with an inductor forming a self-oscillating circuit without needing external voltage sources. We model the nonlinear oscillator dynamics using experimentally measured device characteristics and simulate their behavior under varying optical inputs. Simulations performed on the Fashion-MNIST (Modified National Institute of Standards and Technology) dataset demonstrate that the hybrid ORN–CNN architecture maintains high recognition accuracy and achieves 95% training accuracy and 92% testing accuracy comparable to the fully software-defined CNN with 93% test accuracy while dramatically reducing energy consumption to 24 aJ/operation. Different ORN network topologies were investigated, with asymmetric inductive coupling achieving consistently high performance. These results highlight the potential of oscillator-based networks for low-power analog computing and demonstrate the possibility of using nonlinear physical systems for energy-efficient machine learning applications. This work demonstrates a possible path for integrating device-level oscillator dynamics into future neuromorphic and analog computing platforms.