Brain-inspired, interpretable, resonant recurrent neural networks

Researcher Mark A. Kramer has introduced a novel neural network framework called the Resonant Recurrent Network (RRN), detailed in a recent arXiv paper. Unlike traditional artificial neural networks with non-oscillatory nodes, the RRN is explicitly designed with damped, oscillatory dynamics inspired by biological neurons. The key innovation is structuring these oscillations using two history-dependent terms, connecting them to standard RNN formulations, and applying physical constraints from observed brain rhythms to choose oscillator frequencies. This biologically grounded approach aims to create a more efficient and interpretable model.

In practical tests, the optimized RRN demonstrated its capability by accurately classifying handwritten digits (like those in the MNIST dataset) and simulated neuronal spike trains. Crucially, it achieved this performance with a substantially reduced number of trainable parameters compared to equivalent standard RNNs. The research showed that configuring the oscillator frequencies according to a proposed theory for in vivo brain rhythms led to improved classification accuracy over alternative frequency setups. The authors propose that RRNs could serve as efficient, brain-inspired building blocks for tackling complex tasks in both biological modeling and artificial intelligence, offering a path toward models that are both powerful and more transparent in their internal workings.