All-optical computing based on convolutional neural networks

The rapid development of information technology has fueled an ever-increasing demand for ultrafast and ultralow-energy-consumption computing. Existing computing instruments are pre-dominantly electronic processors, which use electrons as carriers possess von Neumann architecture featured by physical separation storage processing. scaling speed is limited not only data transfer between memory processing units, but also RC delay associated with integrated circuits. Moreover, excessive heating due to Ohmic losses becoming a severe bottleneck both power consumption scaling. Using photons promising alternative. Owing the weak third-order optical nonlinearity conventional materials, building photonic chips under traditional been challenge. Here, we report new all-optical framework realize based on convolutional neural networks. device constructed from cascaded silicon Y-shaped waveguides side-coupled waveguide segments termed “weight modulators” enable complete phase amplitude control in each branch. generic concept can be used equation solving, multifunctional logic operations well many other mathematical operations. Multiple functions including transcendental solvers, multifarious gate operators, half-adders were experimentally demonstrated validate performances. time-of-flight light through network structure corresponds time order several picoseconds ultralow energy dozens femtojoules per bit. Our approach further expanded fulfill complex tasks non-von architectures thus paves way on-chip