Extended Kalman filter demonstrated with a reconfigurable analog photonic neural network

We demonstrate a reconfigurable fixed-point photonic neural network (PNN) based on a coherent photonic processor that interfaces with a digital architecture to perform motion tracking on a damped oscillator using the extended Kalman filter (EKF). The PNN performs comparably to a digital floating-point neural network in two demonstrations; motion tracking of a damped oscillator and voltage monitoring of an RC circuit with nonlinear capacitance. Though the photonic implementation introduces quantization errors, the results of two models over ∼500 time steps do not accumulate tracking errors, improving the uncertainty of the noisy observations by more than 4x in both applications.