Polaritonic Neuromorphic Computing

Artificial intelligence, now finding place in almost all scientific disciplines, is concerned with devices mimicking human cognitive functions to solve certain cumbersome computational tasks. Examples are image- and speech recognition tasks relying on recurrent- or convoluted neural networks. These possess analogy to physical dynamics which has recently given rise to neural inspired hardware in solid state photonic systems. Of interest, the regime of strong light-matter coupling was proposed as a new hardware candidate using networks of microcavity exciton-polariton condensates with promising results. The advantage of exciton-polaritons are their strong nonlinear nature, picosecond timescales, easy write-in and read-out, and malleable optical manipulation. Neural networks must undergo training, which algorithmically tunes the network weights (neuron connections) to achieve best performance. Conventional physical systems cannot do this as their engineering fixes the weights. For exciton-polaritons however, their potential landscape can be designed using optical lasers in a reprogrammable manner with no irreversible engineering performed to the system. Training is then realized by only tuning the optical excitation pattern. The proposal aims at studying networks of exciton-polariton condensates as a new design of artificial intelligence hardware aimed at heuristically solving complex problems, fast parallel processing of large datasets, and time-series prediction tasks.