Computation, Machine Learning & Light

The growing energy consumption and the resulting increase in CO₂ emissions of AI significantly impacts our climate.  Optical computing approaches have made great strides towards the goal of energy-efficient computing necessary for modern deep learning and AI applications. 

Read-in and read-out of data, limit the overall performance of existing approaches. We are therefore working on a framework for multilayer optoelectronic computing that alternates between optical and optoelectronic layers to implement matrix-vector multiplications and rectified linear functions. Our framework is designed for real-time, parallelized operations, leveraging 2D arrays of LEDs and photodetectors connected via independent analog electronics. We are developing custom hardware to implement multiple layers of a deep neural network simultaneously. Our approach significantly reduces the number of read-ins and read-outs required and paves the way for scalable optical accelerators requiring ultra-low energy. We are further developing this technology and expect to beat current state-of-the-art systems for specific applications.

“Low-power scalable multilayer optoelectronic neural networks enabled with incoherent light”, A Song, SN Murty Kottapalli, R Goyal, B Schölkopf, P Fischer, Nature Communications 15 (1), 10692, (2024).