Summary: UCLA researchers in the Department of Electrical and Computer Engineering have developed and experimentally validated a comprehensive training scheme that greatly enhances the robustness of diffractive networks to 3D misalignments and fabrication tolerances in their physical implementation.
Background: As an optical machine learning framework, Diffractive Deep Neural Networks (D2NN) take advantage of data-driven training methods used in deep learning to devise light-matter interaction in 3D for performing a desired statistical inference task. The multi-layer structure of diffractive networks offers significant advantages in terms of their diffraction efficiency, inference capability and optical signal contrast. However, the use of multiple diffractive layers also introduces fabrication and implementation challenges due to the need for structured alignment of multiple layers, reducing optical inference accuracy. To overcome these limitations in the state of the art, a new method to overcome fabrication deficiencies is greatly needed to expand the practical use of D2NNs in machine learning applications.
Innovation: UCLA researchers in the Department of Electrical and Computer Engineering have developed a comprehensive training method that substantially increases the robustness of diffractive optical networks against physical misalignments and fabrication tolerances. This diffractive network design, termed vaccinated D2NN (v-D2NN), models undesired layer-to-layer misalignments during the fabrication process, ensuring that the diffractive networks are trained to maintain their inference accuracies. In addition, the improved training scheme can also be adopted to mitigate other error sources in diffractive network models such as detection noise, fabrication imperfections or artefacts. Importantly, this corrective technology could enable practical machine vision and sensing applications by mitigating various sources of error between the training forward models and the corresponding physical hardware implementations. The utility of this innovation can be extended to a wide array of applications, including advanced biomedical image analysis, speech recognition, holography, and nanophotonic design.
Potential Applications: • 3D imaging • Machine vision • Sensors • Medical imaging • Autonomous vehicle navigation
Advantages: • Mitigated misalignment errors • Reduced detection noise • Resistant to fabrication imperfections and artifacts • Optimized design process
Development to Date: Prototype demonstrated and tested.
Related Papers: Mengu, Deniz, Zhao, Yifan, Yardimci, Nezih T., Rivenson, Yair, Jarrahi, Mona and Ozcan, Aydogan. "Misalignment resilient diffractive optical networks" Nanophotonics, vol. 9, no. 13, 2020, pp. 4207-4219. https://doi.org/10.1515/nanoph-2020-0291
Patent Application:
Misalignment-resilient diffractive optical neural networks
Reference: UCLA Case No. 2020-885
Lead Inventor: Aydogan Ozcan