Description: Long term space missions can greatly benefit from neuromorphic processors enabling in-situ learning in the extreme environments in space. Photonic Tensor Cores provide extremely energy efficient and robust hardware for the computations required for neuromorphic processing. However, current technologies require the significant expenditure of energy, through resistive heaters or current injection diodes, to maintain the state of the neural network for inference. Here we describe a photonic accelerator technology that utilizes the learning from FLASH memory devices to enable non-volatile, 0 energy state retention for neuromorphic processors, enabling the possibility of reaching incredible inference efficiency of 1 femto-joule per operation – a 3 order of magnitude improvement over todays GPUs.

Benefits: Photonic computation, as implemented in optical neural networks, offers the promise to deploy neuromorphic processing with extreme speed and power efficiency. These attributes unlock opportunities for pushing computation/machine learning into smaller systems and spacecraft at the operational edge. On-board Image processing for small craft and real-time imagery inference for autonomous landing are attractive possibilities. Other applications include telescope motion control, terrain classification, medical monitoring and adaptive flight control.

Non-NASA Applications span an even wider gamut, but follow a similar theme: equipping resource constrained platforms with computation. The following list of applications are attractive and large: Commercial space (e.g. Cubesat operators like Planet Labs) UAVs operated in military and commercial settings; Autonomous (i.e. self-driving) car market; Industrial robotics that rely on computer vision;