Scalable Free-Space Optical Neural Networks

The transformative impact of deep neural networks (DNNs) in many fields has motivated the development of hardware accelerators to improve speed and power consumption. We present a novel photonic approach based on homodyne detection where inputs and weights are encoded optically and can be reprogrammed and trained on the fly. This architecture is naturally adapted to free-space optics where both fully-connected and convolutional networks can be implemented and scaled to millions of neurons. By utilizing passive optical fan-out and performing arithmetic coherently with optical interference, this scheme circumvents fundamental limits of irreversible electronic processing. We study the effect of detector shot noise on neural-network accuracy to establish a “standard quantum limit” for this system. This bound, which can be as low as 50 zJ/FLOP, suggests performance below the Landauer (thermodynamic) limit is theoretically possible with photonics.