Felix Effenberger, Pedro Carvalho, Igor Dubinin, and Wolf Singer
PNAS 122 (4) e2412830122
Neocortical circuits are characterized by complex oscillatory dynamics. Whether these oscillations serve computations or are an epiphenomenon is still debated. To answer this question, we designed a computational model of a recurrent network that allows control of oscillatory dynamics (harmonic oscillator recurrent network, HORN). When operating in an oscillatory regime, HORNs outperform nonoscillatory recurrent networks in terms of learning speed, noise tolerance, and parameter efficiency. Moreover, they closely replicate the dynamics of neuronal systems, suggesting that biological neural networks are likely to also exploit the unique properties offered by oscillatory dynamics for computing. The interference patterns provided by wave-based responses allow for a holistic representation and highly parallel encoding of both spatial and temporal relations among stimulus features.
Read the full article at: www.pnas.org