Coincidence detection is all you need.

Celestine Preetham Lawrence

Abstract: This paper demonstrates that the performance of coincidence detection - a classic neuromorphic signal processing method found in Rosenblatt's perceptrons with distributed transmission times, can be competitive to a state-of-the-art deep learning method for pattern recognition. Hence, we cannot remain comfortably numb to the prevailing dogma that efficient matrix-vector operations is all we need; but should enquire with greater vigour if more advanced continual learning methods (running on spiking neural network hardware with neuromodulatory mechanisms at multiple timescales) can beat the accuracy of task-specific deep learning methods. With regards to deployability, coincidence detection is an interpretable shallow learning method and its applications provide a commercial use-case for neuromorphic hardware such as Intel Loihi.