Predictive coding in balanced neural networks with noise, chaos and delays

  • 2020-06-25 05:03:27
  • Jonathan Kadmon, Jonathan Timcheck, Surya Ganguli
  • 25


Biological neural networks face a formidable task: performing reliablecomputations in the face of intrinsic stochasticity in individual neurons,imprecisely specified synaptic connectivity, and nonnegligible delays insynaptic transmission. A common approach to combatting such biologicalheterogeneity involves averaging over large redundant networks of $N$ neuronsresulting in coding errors that decrease classically as $1/\sqrt{N}$. Recentwork demonstrated a novel mechanism whereby recurrent spiking networks couldefficiently encode dynamic stimuli, achieving a superclassical scaling in whichcoding errors decrease as $1/N$. This specific mechanism involved two keyideas: predictive coding, and a tight balance, or cancellation between strongfeedforward inputs and strong recurrent feedback. However, the theoreticalprinciples governing the efficacy of balanced predictive coding and itsrobustness to noise, synaptic weight heterogeneity and communication delaysremain poorly understood. To discover such principles, we introduce ananalytically tractable model of balanced predictive coding, in which the degreeof balance and the degree of weight disorder can be dissociated unlike inprevious balanced network models, and we develop a mean field theory of codingaccuracy. Overall, our work provides and solves a general theoretical frameworkfor dissecting the differential contributions neural noise, synaptic disorder,chaos, synaptic delays, and balance to the fidelity of predictive neural codes,reveals the fundamental role that balance plays in achieving superclassicalscaling, and unifies previously disparate models in theoretical neuroscience.


Quick Read (beta)

loading the full paper ...