Sophie Deneve
Dr Sophie Deneve, CR2
Institut des Sciences Cognitives,
67 Boulevard Pinel
69650
tel : +33 (0) 437 911 239
email : deneve@isc.cnrs.fr
Current research topics:
Many perceptual and motor tasks
performed by the central nervous system have successfully been described in a
Bayesian framework. According to these theories, a probability is assigned to
all possible interpretations of the available sensory and motor information, on
the basis of sensory or motor noise and priors. The percept, or the motor
output, corresponds to the interpretation that is the most likely or has the
most desirable outcome.
I am interested in how neurons and population of neurons compute efficiently in
the presence of noisy and ambiguous sensory input. More specifically, I study
how probabilistic knowledge and uncertainties are represented by cortical
neurons and how recurrent biological neural networks perform, or approximate, bayesian inference.
Optimal
multi-cue integration.
Most sensory and motor variables are represented by the noisy responses of a population neurons tuned to these variables, a form of coding called population code. We show that iterative basis function networks can combine information from several noisy population codes optimally. That is, they perform Bayesian cue integration, a long as the uncertainty associated with each sensory and motor modality is represented by the gain of neural responses in this modality. We apply this approach, in particular, to multi-sensory integration and predict the response properties of neurons in SC, parietal and premotor areas. We are currently testing the predictions of these models through psychophysical experiments and neurophysiological recordings in multisensory brain areas.
Optimal
sensorimotor integration and motor control.
During self-motion, an efficient on-line control of motor effectors require the combination of a forward model, which predicts the sensory consequences of movement from the efferent motor commands, and an inverse model, which compute the motor commands from the desired sensory consequences. We propose that sensorimotor and motor areas contain iterative basis function networks that learn to implement simultaneously an inverse and a forward model. We plan on testing the predictions of these models through psychophysical experiments and neurophysiological recordings in sensorimotor and motor areas.
Bayesian
inference in networks of spiking neurons.
Using a more microscopic and mechanistic approach, we reinterpret synaptic integration in single neurons, neural firing and connectivity in cortical columns as computing and propagating probabilistic evidence in a corresponding Bayesian network. This new theory of neural coding provides computational interpretations of leaky integrate-and fire neurons, balanced excitation and inhibition, poisson-like firing statistics, lateral and feedback connections, synaptic and spike adaptation. Complex statistical models could be learned from the sensory input using simple local rules, such as spike-time dependant plasticity rules.
Publications
Research papers
Deneve, S. and Pouget, A.
Bayesian multisensory and cross-modal spatial links. Journal of Neurophysiology (
Latham, P., Deneve, S. and Pouget,
A. Optimal computations with attractor networks.
Journal of Neurophysiology (
Deneve, S. and Pouget, A. Basis functions for object-centered representations. Neuron . Jan 23;37(2):347-59. 2003.
Pouget, A., Deneve, S and Duhamel, J.R. A Computational Perspective on the Neural Basis of Multisensory Spatial Representations. Nature Review Neuroscience. 3:741-747. 2002.
Deneve, S., Latham, P.E. and Pouget, A. Efficient computation and cue integration with noisy population codes. Nature Neuroscience. 4(8). 2001.
Pouget, A., Deneve, S. and Sejnowski, T.J. Frames of reference in hemineglect:
a computational approach. Progress in Brain Research. 1999; 121:81-97.
Deneve, S., Latham, P.E. and Pouget, A. Reading
population codes: a neural implementation of ideal observers. Nature Neuroscience. 2(8):740-745. 1999.
Pouget, A., Zhang, K., Deneve, S. and Latham,
P.E. Statistically efficient estimation using
population code. Neural Computation, 10:373-401. 1998.
Pouget, A., Deneve, S., Ducom,
J.C. and Latham, P.E. Narrow vs wide tuning curves:
what's best for a population code? Neural Computation.
11:85-90. 1998.
Published conference proceedings
Deneve, S. and Pouget, A. Neural Basis of
Object-Centered Representations.
Deneve,
Deneve, S., Duhamel, J.R., Pouget,
A. A new model of spatial representations in multimodal brain
areas. Advances in Neural Information Processing
Systems. 11. 2001.
Deneve, S., Bayesian inference in recurrent
spiking networks. To appear in Neural Information Processing
Systems. 13. 2004.
Book Chapters
Pouget, A., Deneve, S. and
Sejnowski, T.J. Frames of reference in hemineglect: a computational approach. In
``Neural modeling of brain disorders''. Progress in
Brain Research. 121:81-97. Jim Reggia, Eytan Ruppin and Dennis Glanzman (eds).
Elsevier. 1999.
Pouget, A. Zhang, K., Deneve, S., and Latham,
P.E. Statistically efficient estimation using
population code. In ``Population Coding''.
Abbott, L. and Sejnowski, T.J. (eds). MIT Press. 1999.
Pouget, A., Deneve, S., and Latham, P.E. Fisher
Information: Relevance for Theories of Cortical Computation and Visual Attention.
In ``Visual Attention and Neural Circuits''. Braun,
J., Koch. C. and Davis, J. (eds).
2000.
Pouget, A., Deneve, S. and Duhamel,
J.R. A neural theory of multisensory
spatial representation. In “Crossmodal
Space and Crossmodal attention” Spence C. and Driver,
J. (Eds)
Pouget, A., Deneve, S. and Duhamel,
J.R. A neural theory of multisensory
spatial representation. In “Crossmodal Space
and Crossmodal attention” Spence C. and Driver, J. (Eds)