The oscillatory retinal neuron network does not require an external voltage source, has energy costs thousands of times lower than current approaches, and exhibits performance comparable to state-of-the-art GPU-based convolutional neural networks.
Among artificial intelligence approaches, convolutional neural networks (CNNs) are still known to be power-hungry. This is primarily due to bottlenecks that occur when moving data between memory and computing units. New research using analog coupled oscillators aims to provide an energy-efficient alternative.
Abbasi Jalal et al. developed an approach that replaces the first layer of a CNN with a network of oscillatory retinal neurons (ORNs). By using a 3×3 network of photodetectors with a photoactivated negative differential resistance device coupled to an inductor, the self-oscillating circuit requires no voltage from an external power source.
This hybrid device demonstrates the potential of nonlinear physical systems in energy-efficient machine learning applications.
“The default assumption is that the convolutional layer is an expensive layer and we use better digital multipliers or memristor crossbars to accelerate it,” said author Seyedeh Atiyeh Abbasi Jalal. “We can replace this with the inherent physics of a coupled oscillator network, and claim that the energy cost per operation is about six orders of magnitude lower than today’s state-of-the-art graphics processing units.”
The group experimentally measured the properties of the device and used it to model its dynamics and generate Fourier spectra for comparison with more traditional CNNs.
This network achieved a test accuracy of 92% compared to 93% for a full software CNN. The ORN contributes significantly to the work done by the hybrid CNN, consuming approximately 24 atjoules per operation. After testing 30 different bonding topologies, the group found that asymmetric inductive bonding was the most stable.
The group is looking to build ORN arrays with readouts, amplifiers, and other components to evaluate performance under real-world noise and device variations.
sauce: “Integrating oscillator-based feature extraction for energy-efficient convolutional neural networks” by Seyedeh Atiyeh Abbasi Jalal, Ragib Ahsan, Zezhi Wu, Mirbehrad Mousavi, and Rehan R. Kapadia; applied physics journal (2026). This article can be accessed from: https://doi.org/10.1063/5.0323116 .
