Fig. 1: Schematic of the optical coherent dot-product chip (OCDC).

a Simplified model of a neural network for regression. A neural network is typically composed of nodes, connections, and activation functions. Three typical functions (sigmoid, rectified linear unit (ReLU), and hyperbolic tangent (tanh)) are depicted in the figure. Positive, negative, and zeros nodes are depicted with different colors. For fully connected networks, a node represents a single value and a connection is a weight. For convolutional networks, a node is a feature map and a connection is a convolutional kernel. b Histograms of weights in LSTM³¹, U-net³², and AUTOMAP³⁰, respectively. They approximately obey normal distribution with a mean value of zero. c Histograms of activated values in LSTM, U-net, and AUTOMAP, respectively. The distribution depends on the activation functions being used. d The conceptual schematic of the OCDC. It contains several parallel branches for dot product and one extra branch for coherent detection. The optical field in each branch is symbolized with red curves. The push-pull configured modulator imposes amplitude-only modulation to the optical field without introducing phase shift. Hence, the phase of each branch is stable