2019/05/15 by Mario Miscuglio, Gina C. Adam, Miscuglio, Mario +5
Computer Science · Engineering · #Applied Physics (physics.app-ph) #Disordered Systems and Neural Networks (cond-mat.dis-nn) #FOS: Physical sciences #Materials Science (cond-mat.mtrl-sci) #Neural Networks and Reservoir Computing #Optical Network Technologies #Photonic and Optical Devices
paper · pdf · doi:10.48550/arxiv.1905.06371
openalex publication_date 2019/05/15 · openalex created_date 2022/07/23 · openalex updated_date 2026/07/28
Driven by machine-learning tasks neural networks have demonstrated useful\ncapabilities as nonlinear hypothesis classifiers. The underlying technologies\nperforming the dot product multiplication, the summation, and the nonlinear\nthresholding on the input data in electronics, however, are limited by the same\ncapacitive challenges known from electronic integrated circuits. The optical\ndomain, in contrast, provides low delay interconnectivity suitable for such\nnode distributed non Von Neumann architectures relying on dense node to node\ncommunication. Thus, once the neural network's weights are set, the delay of\nthe network is just given by the time of flight of the photon, which is in the\npicosecond range for photonic integrated circuits. However, the functionality\nof memory for storing the trained weights does not exists in optics, thus\ndemanding a fresh look to explore synergies between photonics and electronics\nin neural networks. Here we provide a roadmap to pave the way for emerging\nhybridized photonic electronic neural networks by taking a detailed look into a\nsingle node's perceptron, discussing how it can be realized in hybrid photonic\nelectronic heterogeneous technologies. We show that a set of materials exist\nthat exploit synergies with respect to a number of constrains including\nelectronic contacts, memory functionality, electrooptic modulation, optical\nnonlinearity, and device packaging. We find that the material ITO, in\nparticular, could provide a viable path for both the perceptron weights and the\nnonlinear activation function, while simultaneously being a foundry process\nnear material. We finally identify a number of challenges that, if solved,\ncould accelerate the adoption of such heterogeneous integration strategies of\nemerging memory materials into integrated photonics platforms for real time\nresponsive neural networks.\n