Mirko Goldmann
The field of photonic computing is advancing and recent hardware developments achieve competitive benchmarks in terms of accuracy, energy efficiency, inference speed, and integration size. Many of these hardware substrates are built around the reservoir computing (RC) framework, where only a linear output layer is optimized to solve the task. Whereas RC is a practical approach to train physical hardware, performance boosts of these setups can mainly increase with the size and complexity of the physical setup. As only the output layer and a small amount of coarse-grained parameters are trained, the hardware is not perfectly optimized for the given task. This is in contrast to deep learning concepts that show enhancing abilities with the optimization of increasing amounts of fine-grained parameters. While recent work focuses on the training of hardware neural networks (NN), many of them still have flaws such as relying on additional digital computing, and showing a limited scalability. In this talk, I discuss what can be expected when optimizing NN beyond the RC paradigm. I will then review recent advances in learning algorithms that allow the training of neuromorphic/physical systems. I will also elaborate on two showcases of using sampling-based evolutionary algorithms to train a photonic neural network based on a large area vertical cavity surface emitting laser and a photonic integrated circuit.
https://zoom.us/j/98286706234?pwd=bm1JUFVYcTJkaVl1VU55L0FiWDRIUT09
Contact details:
Miguel C. Soriano Contact form