Attojoule Modulators for Photonic Neuromorphic Computing

Neuromorphic networks are computational algorithms and network models inspired by signal processing in the brain with societal-relevant applications in machine-learning such as speech and image recognition, non-linear... [ view full abstract ]

Neuromorphic networks are computational algorithms and network models inspired by signal processing in the brain with societal-relevant applications in machine-learning such as speech and image recognition, non-linear optimization, and real-time simulation. Recent progress by both industry and academia has demonstrated compute efficiencies surpassing the digital efficiency wall set by von-Neumann compute architectures. An artificial neural network consists of a set of input artificial neurons connected to both the hidden- and the output layer. Within each layer, information propagates by a linear combination such as matrix multiplication followed by the application of a nonlinear activation function. Optical neural network (ONN) implementations offer unique advantages over microelectronics because the linear transformation (e.g. matrix multiplication) can be executed in the optically with temporal response times <10ps.

Neural networks require both a weighting of inputs and a nonlinear activation function operating on their sum. Neural network weighting has been demonstrated in integrated photonics with both interferometric and ring-based wavelength division multiplexing. While direct nonlinearity in optics is difficult to achieve without high optical powers, an electro-optic nonlinearity can be created by directly coupling a photodiode to electro-optic modulator. The low capacitance of directly coupling the components results in operating speeds >10GHz with relatively low power consumption. Here we present a closed form equation for the activation functions created by graphene and quantum well electro-optic absorption modulators capacitive coupled to photodiodes. Our modulator-geometry based and thermal-noise analysis shows that such electro-optic neurons produce SNRs around 60. Performing an MNIST classification inference test on an AlexNet-based neural network with these electro-optic nodes, with accuracies of about 95% starting a laser power level around 3mW and 10mW for the QW and Graphene- based modulator, respectively. Our findings show regions of realistic operating performance of future optical and photonic neural networks using electro-optic analogue (non-spiking) neurons.

We report on using the transfer function of electrooptic absorption modulators as nonlinear activation functions of photonic neurons and show 95% accuracy of MNIST classification inference on an AlexNet in optical artificial neural networks.