Abstract

Photonic analogue implementations of neural networks represent a novel computational paradigm beyond classical Turing/Von Neumann architectures, which demonstrates extremely powerful performance for non-algorithmic problems. Similarly to deep learning schemes [1], efficient operation of analogue neural networks requires large scale complex network architectures, just as schematically illustrated in Fig. 1(a). Nevertheless, in contrast to digital systems, analogue implementations will always be prone to noise of different origin. Therefore, understanding the role of noise and its propagation through the multiple network layers is essential for all connectionist computing schemes. However, no systematic studies in this direction have so far been reported yet. We analyze the impact of noise on the performance of analogue and large-scale neural networks implemented in photonic hardware.

© 2019 IEEE