Identify the roles of recurrent connections and other factors in RSNN decoding performance

Determine the specific contributions of individual factors—particularly recurrent connections—within the bigRSNN and tinyRSNN architectures to the observed improvements in decoding macaque finger velocities from cortical spike trains, by isolating and quantifying how each factor impacts decoding accuracy.

Background

The paper introduces two recurrent spiking neural network decoders—bigRSNN and tinyRSNN—that outperform existing feed-forward SNN and ANN baselines on finger-velocity decoding from macaque cortical spike trains while meeting resource constraints. Despite these gains, the authors note that the mechanisms underlying the improvements are not yet understood.

Specifically, they highlight uncertainty about which architectural or training components, with emphasis on recurrent connections, drive the observed performance increases. They point to the need for a detailed analysis to disentangle and quantify the impact of these factors.