Resonance for analog recurrent neural network

Wave-based analog computing enjoys benefits of intrinsic parallelism and it can be extremely energy efficient compared to digital computing [1]. However, the transient nature of propagating waves makes it difficult to construct memory in the wave domain. Since memory is indispensable for computing in the temporal data, researchers have to resort to other means to realize the effect of memory such as optoelectronic conversion, routing through long waveguides and random internal feedback. In all these works, the memory is implicitly built into the complex structures, and physical intuition and interpretation are lacking. However, in the resonance system, we can include resonators with different lifetimes to realize both short-term and long-term memory. Here through a set of theoretical work, we propose resonance as a general form of memory to be used for complex temporal computing and advanced recurrent models such as LSTM. The findings here have broad impact and help to shape the future computing based on optical and acoustic waves.

[1] Wetzstein, Gordon, et al. "Inference in artificial intelligence with deep optics and photonics." Nature 588.7836 (2020): 39-47.