Related Experiment Video
Updated: Aug 6, 2026
06:36
Optrode Array for Simultaneous Optogenetic Modulation and Electrical Neural Recording
Published on: September 1, 2022
Wavelength-encoded neuromorphic inference enabled by microcavity MoS2 photodetector arrays
Xiang Chen1, Kangjian Di2,3, Fuhao Yu3,4
1School of Electronic Science and Engineering, National Laboratory of Solid-State Microstructures, Nanjing University, Nanjing, China.
Nature Communications
|July 18, 2026
Summary
This study introduces a novel bio-inspired vision system that unifies sensing and computation in each pixel. This approach significantly reduces latency and energy costs for efficient neuromorphic inference.
Area of Science:
- Optoelectronics
- Neuromorphic Engineering
- Materials Science
Background:
- Biological vision systems efficiently integrate spectral sensing and temporal processing, unlike conventional hardware that separates photodetection and computation, leading to high latency and energy use.
- Existing optical neural networks face scalability challenges.
- There is a need for integrated, efficient, and scalable vision hardware.
Purpose of the Study:
- To propose a bio-inspired optoelectronic inference architecture that unifies sensing, weighting, and accumulation within each pixel.
- To leverage spectral domain for encoding neural network weights and material properties for temporal accumulation.
- To develop a compact and scalable solution for in-sensor neuromorphic inference.
Main Methods:
- Developed a Fabry-Perot microcavity-integrated MoS2 photodetector array.
- Engineered cavity wavelength selectivity to encode neural network weights spectrally.
- Utilized the finite carrier lifetime of MoS2 for analog temporal accumulation without external memory.
- Implemented post-training optical Hessian pruning for complexity reduction.
Main Results:
- Achieved high test accuracies: 99.6% on MNIST, 94.8% on CIFAR-10, and 94.0% on the Free Spoken Digit Dataset.
- Demonstrated unified sensing, weighting, and accumulation within each pixel.
- Showcased robustness and reduced optical complexity through pruning.
Conclusions:
- The proposed architecture offers a compact and efficient route toward wavelength-aware, in-sensor neuromorphic inference.
- This bio-inspired approach overcomes limitations of conventional optoelectronic vision systems.
- The integration of spectral encoding and temporal accumulation in a single pixel represents a significant advancement in neuromorphic hardware.
