Neural Inference Group


Research center

45 rue d’Ulm
75230 Paris
Marc Mézard


ED 3C - 158
Université Pierre et Marie Curie


Laboratoire de Neurosciences Cognitives
U 960
PSL, Labex de l’institut d’Etudes cognitives de l’ENS


Constructing precisely computing networks with biophysical spiking neurons. MA Schwemmer#, AL Fairhall, S Denéve, ETShea-Brown. J Neurosci 2014

What visual illusions teach us about schizophrenia. CE Notredame#, D Pins, S Deneve, R Jardri Frontiers in integrative neuroscience 8 2014

Neural oscillations as a signature of efficient coding in the presence of synaptic delays. Chalk M#, Gutkin B, Denève S. Elife. 2016 Jul 7;5. pii: e13824. doi: 10.7554/eLife.13824.

Are Hallucinations Due to an Imbalance Between Excitatory and Inhibitory Influences on the Brain? Jardri R#, Hugdahl K, Hughes M, Brunelin J, Waters F, Alderson-Day B, Smailes D, Sterzer P, Corlett PR, Leptourgos P, Debbané M, Cachia A, Denève S. Schizophr Bull. 2016 Sep;42(5):1124-34. doi:10.1093/schbul/sbw075.

Efficiency turns the table on neural encoding, decoding and noise. Deneve S, Chalk M#. Curr Opin Neurobiol. 2016 Apr;37:141-8. doi: 10.1016/j.conb.2016.03.002. Review.

Efficient codes and balanced networks. Denève S, Machens CK. Nat Neurosci. 2016 Mar;19(3):375-82. doi: 10.1038/nn.4243. Review.

Fields of research

Computational neurosciences / neural theory

Research Theme

Currently developping several research topics

1- Spike-based predictive coding:

Progress in understanding brain functions rely in great part on filling the conceptual and experimental gaps between different levels ofanalysis, from single neurons to behaviur. Thus, in “rate models”, single spikes are as meaningless as single molecules in a gas areto thermodynamics laws, while function and behaviour emerge from the responses of very large networks.My research aims at developing an alternative approach, spike-based predictive coding, relating spiking neural activity and dynamicsdirectly to their functional interpretation. It combines two relatively straightforward hypotheses: Neural networks reliably estimate thestate of the environment or body based on their inputs and prior experience (optimal inference). And their dynamics insures that theseestimates can be decoded from their spike trains by postsynaptic integration (self consistency). By monitoring and decoding its ownoutputs, the neural structure itself closes the loop between computation and dynamics, descriptive and functional interpretations ofneural activity.

2-Normal and pathological inference in brain circuits

Recent years have seen the growing use of Bayesian models to describe behavior, perception and reasoning. These modelsformalize problems in sensory perception, motor control or behavioral strategies as probabilistic inference and learning in causalmodels. The goal of this research is to adapt this powerful set of tools to further our understanding of the function and dynamics ofbiological neural networks, and how they underlie cognition. Our particular focus is the neural implementations of hierarchical causalinference. We identified building blocks of neural inference, in the form of elementary probabilistic computations such as integration ofnoisy sensory evidence (bottom up processing), modulations by prior beliefs and feedback from higher order representations (topdown processing), and competitions between alternative interpretations of a sensory scene (explaining away). Recently we appliedthese models of hierarchical neural inference to positive syndromes of schizophrenia. We interpreted hallucinations and delusions asimpaired message passing in hierarchical beliefs networks (and corresponding neural networks). We investigated potential candidateneural mechanisms, in particular a possible dysfunction in the inhibitory loops normally preventing top-down predictions from beinginterpreted as new sensory observations.