Meta-neural-network for Realtime and Passive Deep-learning-based Object Recognition
This work addresses the trade-off between device size, cost, and energy consumption versus recognition speed and accuracy for applications in acoustics, optics, and related fields, though it is incremental as it builds on existing deep-learning and metamaterial concepts.
The authors tackled the problem of real-time object recognition by proposing a passive meta-neural-network that uses acoustic scattering to recognize complex objects, demonstrating its capability through handwritten digit recognition with a compact design and high-resolution performance. They achieved this by mimicking standard neural networks with metamaterial unit cells, enabling deep-subwavelength phase shifts as learnable parameters.
Deep-learning recently show great success across disciplines yet conventionally require time-consuming computer processing or bulky-sized diffractive elements. Here we theoretically propose and experimentally demonstrate a purely-passive "meta-neural-network" with compactness and high-resolution for real-time recognizing complicated objects by analyzing acoustic scattering. We prove our meta-neural-network mimics standard neural network despite its small footprint, thanks to unique capability of its metamaterial unit cells, dubbed "meta-neurons", to produce deep-subwavelength-distribution of discrete phase shift as learnable parameters during training. The resulting device exhibits the "intelligence" to perform desired tasks with potential to address the current trade-off between reducing device's size, cost and energy consumption and increasing recognition speed and accuracy, showcased by an example of handwritten digit recognition. Our mechanism opens the route to new metamaterial-based deep-learning paradigms and enable conceptual devices such as smart transducers automatically analyzing signals, with far-reaching implications for acoustics, optics and related fields.