Optics, Photonics & Laser

Biography

Biography: Alessandro Bile

Abstract

Neuromorphic models [1,2] are proving capable of performing complex machine learning tasks, overcoming the structural limitations imposed by software systems [3] and electronic neuromorphic models [4]. Recently a photon solitonic neuron model [5] has been developed that is able to receive information, process it and store it. The work we present creates a soliton neural network (SNN) through the interfacing of these neurons [6]. The networks consist of a succession of X-shaped junctions which, recognizing the information propagated within the guides, switch by identifying specific preferential trajectories which constitute bit-by-bit memories. The network can memorize and subsequently use the acquired information to recognize further unknown information. The peculiar characteristic of the SNN is its ability to learn in a plastic way, similarly to what happens in the biological tissue. In the nervous system, neurons exchange signals and recognize incoming patterns thanks to the creation, consolidation and destruction of synaptic bridges. Similarly, our neurons can save the information they receive by self-modifying their structures through variations in the refractive index. We propose a neuromorphic model based on a solitonic-waveguide X-junctions interfacing, as shown in fig. 1, obtained by interfacing two input soliton neurons (fig 1a), whose channels give rise to three layers: an input layer, a hidden layer and an output layer as reported in fig 1b. Fig 2. re-proposes network training on four different configurations. Each configuration is characterized by only one channel active while the others are off. Training consists in injecting for several cycles the signal that modifies the refractive index map. Then, inserting signals into all the input channels show a high output response only in correspondence with the trained channel.