Existence and properties of LPUs in spiking recurrent neural networks

Determine whether spiking recurrent neural networks can instantiate self-sufficient and universal latent dynamical systems (latent processing units) and, if so, evaluate whether such LPUs can achieve generality and explanatory power comparable to those in firing-rate architectures under spike-based communication constraints.

Background

The paper’s main results are developed for firing-rate style recurrent networks with linear encoding and nonlinear embedding. Extending these ideas to spiking recurrent neural networks (sRNNs) would increase biological realism but raises nontrivial questions about whether LPUs can emerge and maintain universal computational properties in discrete spike-based communication regimes.

Clarifying existence and performance of LPUs in sRNNs would help bridge theory with biological circuit implementations and impact applications such as robust brain–machine interfaces.