Photonic reservoir computing with SOAs and delays

Reservoir Computing is a methodology coming from the field of machine learning and neural networks that has been successfully used in pattern classification problems. Instead of feeding inputs directly into a linear classifier, the classifier takes its input from a reservoir with recurrence where the inputs have been mixed. One classical implementation employs a recurrent neural network with hyperbolic tangent functions in the nodes as a reservoir. On an isolated digit speech recognition task, with 3 dB SNR babble noise added [2], these tanh reservoirs achieve a performance around 7%. In a previous paper we have already shown that a network of Semiconductor Optical Amplifiers (SOA) can be used as a reservoir on a simple signal classification task, making it an interesting hardware implementation [1]. Here, such an SOA network will be used for speech recognition.