Greedy Learning in a Large Scale Photonic Network

We have recently succeeded in the implementation of a large scale recurrent photonic neural network hosting up to 2025 photonic neurons. All network internal and readout connections are physically implemented with fully parallel technology. Based on a digital micro-mirror array, we can train the Boolean readout weights using a greedy version of greedy learning. We find that the learning excellently converges. Furthermore, it appears to possess a conveniently convex-like cost-function and demonstrates exceptional scalability of the learning effort with system size. I will introduce our photonic neural network in detail and give a general motivation of photonic systems for neural network processors. Finally, I will discuss the obtained findings of the learning procedure in light of their relevance for hardware implemented neural networks.