Automatic classification of video using a scalable photonic neuro-inspired architecture

We propose a physical alternative of software based approaches for advanced classification task by considering a photonic-based architecture implementing a recurrent neural network with up to 16,384 physical neurons. This architecture is realized with o.-the-shelf components and can be scaled up to hundred thousand or millions of nodes while ensuring data-efficient training strategy thanks to the reservoir computing framework. We use this architecture to perform a challenging computer vision task: the classification of human actions from a video feed. For this task, we show for the first time that a physical architecture with a simple learning strategy, consisting of training one linear readout for each class, can achieve a > 90% success rate in terms of classification accuracy. This rivals the deep-learning approaches in terms of level of performance and hence could pave the way towards novel paradigm for efficient real-time video processing at the physical layer using photonic systems.