Dynamics of random recurrent networks with correlated low-rank structure
Friedrich Schuessler, Alexis Dubreuil, Francesca Mastrogiuseppe, Srdjan Ostojic, Omri Barak
Abstract
Learning in the brain happens on the basis of pre-existing, task-unrelated connectivity, and structural components created during learning are correlated to this initial connectivity. To investigate how pre-existing and learnt connectivity interact, the authors study dynamics in nonlinear neural network models where connectivity consists of a random part and a correlated low-rank perturbation. By computing fixed points and their stability, they show how correlations between pre-existing and learnt connectivity enrich the dynamical repertoire of the model.
Topics & Concepts
Dynamics (music)Computer scienceStatistical physicsNonlinear systemArtificial neural networkArtificial intelligenceBasis (linear algebra)MathematicsRecurrent neural networkNetwork dynamicsPoint processNonlinear dynamical systemsNetwork structureFixed pointDynamical systems theoryTheoretical computer scienceRandom graphComplex networkAlgorithmStability (learning theory)Statistical learningStochastic processCorrelationDeep learningKey (lock)Point (geometry)Neural dynamics and brain functionFunctional Brain Connectivity StudiesStochastic Gradient Optimization Techniques