Beecroft Building
Prof. Demetri Psaltis (EPFL)
Abstract:
The optical implementation of neural networks can be advantageous compared to electronics in terms of power consumption. This derives from the fact that the energy required to transmit information optically can be nearly independent of the distance between the emitter and the receiver. Consequently, optics can be particularly suitable for hardware implementations of neural networks due to the dense connectivity of neural architectures. Neural networks, however, use the strengths of the interconnections between the processing units (the “neurons”) as computing and storage elements. An optical neural network must therefore include a mechanism that allows it to be programmed or trained. In this presentation we will present recent results for programming optical learning machines [1,2,3].
[1] Programming nonlinear propagation for efficient optical learning machines
I Oguz, JL Hsieh, NU Dinc, U Teğin, M Yildirim, C Gigli, C Moser, D Psaltis
Advanced Photonics 6 (1), 016002-016002
[2] Forward–forward training of an optical neural network
I Oguz, J Ke, Q Weng, F Yang, M Yildirim, NU Dinc, JL Hsieh, C Moser, ...
Optics Letters 48 (20), 5249-5252
[3] Nonlinear Processing with Linear Optics
M Yildirim, NU Dinc, I Oguz, D Psaltis, C Moser, arXiv preprint arXiv:2307.08533