Emulation and modelling of semiconductor optical amplifier-based all-optical photonic integrated deep neural network with arbitrary depth

Abstract We experimentally demonstrate the emulation of scaling semiconductor optical amplifier (SOA) based integrated all-optical neural network in terms number input channels and layer cascade, with chromatic at neuron monochromatic output conversion, obtained by exploiting cross-gain-modulation effect. propose a noise model for investigating signal degradation on processing after cascades SOAs, we validate it via experimental results. Both experiments simulations claim that (AON), wavelength conversion as non-linear function, is able to compress noisy inputs. This suggests use SOA-based AON may allow building networks arbitrary depth. In fact, an arbitrarily deep network, built out seven-channel AONs, shown guarantee error minor than 0.1 when operating power levels −20 dBm/channel 6 dB dynamic range. Then results, extended layers, suggest cascading interconnecting multiple these monolithically possible build 12-inputs/neuron 12 neurons/layer depth scaling, or 18-inputs/neuron 18-neurons/layer single implementation, maintain <0.1. Further improvement height scalability can be optimizing power.