Mixed photonic/electronic neural network based on microLED arrays

We present a novel approach for the implementation of a mixed photonic/electronic neural network using gallium nitride based microLEDs as an efficient key element. The system is fully analog, with the synaptic fan-out and weighting implemented in free-space optics, while the fan-in and nonlinear activation functions are realized by optoelectronic components. Activations are represented by patterns of incoherent light generated by microLED arrays. Spatial light modulators represent the weighting matrices in the optical path, and the summation of the weighted activations is realized by photodetectors. We practically demonstrate our approach with a prototype, implementing a single-layer artificial neural network for the classification of the MNIST dataset. A test accuracy of 86.8% is achieved. A key feature of this photonic/electronic implementation is that an efficient conversion between the electrical and optical domains is realized by microLEDs, combining the advantages of the efficiency of synaptic connections through optics and nonlinear activation functions in electronics, which have the potential for miniaturization and integration. We estimate, that an upscaled version of our prototype can achieve an energy efficiency of around 14 fJ per multiply-accumulate operation using state-of-theart microLEDs. Finally we present a  scalability analysis, showing that for a fully connected layer, for which a total of N multiply-accumulate operations have to be performed, in the average case the energy requirement scales with O(√N), while the energy requirement of conventional computers, as well as digital neuromorphic computing approaches, scales with O(N). Therefore, we suggest to follow this approach in order to fully exploit the potential of mixed optical/electrical neuromorphic systems and possibly demonstrate superiority over conventional GPUs.