Ultra-Compact Non-linear Metamaterial Wave Processors for Analog Deep Learning – MetaLearn

Information processing based on tailored wave-matter interactions emerges as an appealing alternative to conventional digital processors in terms of speed and energy consumption, in particular for low-latency low-power Internet-of-Things sensors with artificial intelligence (AI). We will introduce locally resonant metamaterials with non-linear and programmable inclusions as an ideal platform to implement ultra-compact AI wave processors, the compactness being of particular relevance for common long-wavelength signals like microwaves and sound. We will develop a generic semi-analytical model of such wave processors based on coupled dipoles. We will fabricate acoustic prototypes mainly targeted at processing temporally encoded information (e.g. in speech) and radiofrequency prototypes mainly targeted at processing spatially encoded information (e.g. for gesture recognition).