Spike based inference in a network with divisive inhibition

Sophie Denéve (1) , Timm Lochmann (1) , Udo Ernst (2)

1 Group for Neural Theory [Paris]
2 Institute for Theoretical Physics - Bremen

Lateral inhibition has been found in many brain areas related to processing of sensory information. We examine the functional properties of a specific type of such inhibitory connectivity. Our analysis is grounded on a generative model (GM) describing the statistical relation between objects in the world (also called hidden states) and the neural response they evoke. For a set of reasonable assumptions regarding how information is processed in single cells and the assumptions contained in the generative model, we show that divisive inhibition enables a network to approximate optimal inference. We test the goodness of this approximation with numerical simulations. In these, we compare how well a sequence of hidden states can be reconstructed from the network activity. We show how this inference about the states can be formulated in the terminology of Hidden Markov Models (HMMs). This allows us to compare the networks performance to solutions obtained from standard HMM inference algorithms. We find divisive inhibition to clearly improve network performance and the resulting estimates close to the optimal ones. We discuss the results and mention further properties and extensions of our approach.

Mots clés: Artificial neural networks; Probabilistic Models and Methods; Neural coding; Population coding

Domaines: Neurosciences [q-bio.NC]; Neurosciences

Sophie Denéve, Timm Lochmann, Udo Ernst. Spike based inference in a network with divisive inhibition. Deuxième conférence française de Neurosciences Computationnelles, "Neurocomp08", Oct 2008, Marseille, France. ⟨hal-00331564⟩