QuATON: quantization aware training of optical neurons
Optical processors, built with "optical neurons," can perform large-scale high-dimensional linear operations at the speed of light. With the current advances in micro-fabrication, such optical processors can now be 3D-fabricated, but at limited precision, eventually leading to a model mismatch due to quantized optical weights. To address this issue, we propose a quantization-aware training framework. Our approach accounts for physical constraints during the training process, leading to robust designs. We numerically demonstrate that our approach can design state-of-the-art optical processors using diffractive networks for multiple tasks despite quantized learnable parameters. We thus lay the foundation upon which improved optical processors may be 3D-fabricated in the future.
Authors
- Takahiro Nambara
- Alex Matlock (ORCID: https://orcid.org/0000-0002-4337-2715)
- Peter T. C. So (ORCID: https://orcid.org/0000-0003-4698-6488)
- Hasindu Kariyawasam (ORCID: https://orcid.org/0000-0003-3074-7239)
- Yuichiro Kunai
- Quansan Yang (ORCID: https://orcid.org/0000-0001-8029-1327)
- Dushan N. Wadduwage (ORCID: https://orcid.org/0000-0002-0689-9454)
- Ramith Hettiarachchi (ORCID: https://orcid.org/0000-0001-7070-8213)
- Hiroyuki Kusaka (ORCID: https://orcid.org/0000-0002-0326-7179)
- Ed Boyden
Institutions
- University of Washington (US)
- Myanmar Institute of Theology (MM)
- Fujikura (Japan) (JP)
- Moscow Institute of Thermal Technology (RU)
- Hydrobiological Institute (MK)
- Koch Institute for Integrative Cancer Research At MIT (US)
- Massachusetts Institute of Technology (US)
- Fujikura (United States) (US)
Publication Details
- Journal
- Optics Express
- Published
- 2025-12-10
- DOI
- https://doi.org/10.1364/oe.546074
- Primary Topic
- Neural Networks and Reservoir Computing
- Type
- article
- Field-Weighted Citation Impact
- 0.00