Hybrid convolutional neural network yields energy-efficient approach with analog first layer

JUN 18, 2026

Oscillatory retinal neuron networks don’t require external voltage sources and show comparable performance to cutting-edge GPU-based convolutional neural networks, for energy costs thousands of times less than current approaches.

Adam Liebendorfer

DOI: 10.1063/10.0044206

Hybrid convolutional neural network yields energy-efficient approach with analog first layer internal name — Hybrid convolutional neural network yields energy-efficient approach with analog first layer lead image

Among artificial intelligence approaches, convolutional neural networks (CNNs) remain notoriously power intensive. This stems in large part from a bottleneck created when moving data between memory and compute units. New work with analog coupled oscillators looks to provide energy-efficient alternatives.

Abbasi Jalal et al. have developed an approach that replaces the first layer of a CNN with a network of oscillatory retinal neurons (ORNs). Using a 3x3 network of photodetectors with optically activated negative differential resistance devices coupled to an inductor, the self-oscillating circuit doesn’t require voltage from external sources.

The hybrid device demonstrates the potential of nonlinear physical systems for energy-efficient machine learning applications.

“The default assumption is that the convolutional layer is the expensive layer, and you accelerate it with better digital multipliers or with memristor crossbars,” said author Seyedeh Atiyeh Abbasi Jalal. “We’re making the case that you can just replace it with the intrinsic physics of a coupled oscillator network, and the energy cost is around six orders of magnitude lower per operation than today’s cutting-edge graphics processing units.”

The group measured the device’s characteristics experimentally, which were then used to model its dynamics, producing Fourier spectra to compare it to more conventional CNNs.

The network achieved 92% test accuracy, compared to 93% for fully software CNNs. The ORN contributes significantly to the work done by the hybrid CNN, with about 24 attojoules per operation. After testing 30 different coupling topologies, the group found that asymmetric inductive coupling was the most consistent.

The group looks to build an ORN array along with readout, amplifiers and other components to assess its performance under real noise and device variation.

Source: “Integration of 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, Journal of Applied Physics (2026). The article can be accessed at https://doi.org/10.1063/5.0323116 .