Neural network connectivity by optical broadcasting between III-V nanowires.

Biological neural network functionality depends on the vast number of connections between nodes, which can be challenging to implement artificially. One radical solution is to replace physical wiring with broadcasting of signals between the artificial neurons. We explore an implementation of this concept by light emitting/receiving III-V semiconductor nanowire neurons in a quasi-2D waveguide. They broadcast light in anisotropic patterns and specific regions in the nanowires are sensitised to exciting and inhibiting light signals. Weights of connections between nodes can then be tailored using the geometric light absorption/emission patterns. Through detailed simulations, we determine the connection strength based on rotation and separation between the nanowires. Our findings reveal that complex weight distributions are possible, indicating that specific neuron geometric patterns can achieve highly variable connectivity as needed for neural networks. An important design parameter is matching the wavelength to the specific physical dimensions of the network. To demonstrate applicability, we simulate a reservoir neural network using a hexagonal pattern of nanowires with random angular orientations, displaying its ability to perform chaotic time series prediction. The design is compatible with integration on Si substrates and can be extended to other nanophotonic components.