Performance Enhancement via Synaptic Plasticity in an Integrated Photonic Recurrent Neural Network with Phase-Change Materials

Synaptic plasticity, i.e. the ability of synaptic connections to strengthen or weaken depending on their input, is a fundamental component of learning and memory in biological neural networks [1]. This property allows the network parameters to directly adapt to the input signal, thus without being externally tuned by a training algorithm. In contrast with this paradigm, the most popular and successful artificial neural network (ANN) models are nowadays based on backpropagation, which usually requires full observability of the network states and precise parameter tuning. In practice, these requirements strongly limit the scalability of neuromorphic hardware and backpropagation is not considered biologically plausible [2].