DEEPEN SIGNED

Project "DEEPEN" data sheet

The following table provides information about the project.

 Partnership

Take a look of project's partnership.

Map

+-

Leaflet | Map data © OpenStreetMap contributors, CC-BY-SA, Imagery © Mapbox

 Project objective

The principles of sensory perception are still a large experimental and theoretical puzzle. A strong difficulty is that perception emerges from networks of recurrently connected brain areas whose activity and function are poorly approximated by current generic mathematical models. These models also fail to explain many of the fundamental structures effortlessly identified by the brain (shapes, objects, auditory or tactile categories). I here propose to establish a new approach combining high-throughput population recoding methods with a tailored theoretical framework to derive computational principles operating throughout sensory systems and leading to biologically structured perception. This approach follows on the recent mathematical proposal, suggested by Deep Machine Learning methods, that complex perceptual objects emerge through series of simple nonlinear operations combining increasingly complex sensory features along the sensory pathways. Starting with the mouse auditory system as a model pathway, we will recursively extract, with model-free methods, the main nonlinear sensory features encoded in genetically tagged output and local neurons at different processing stages, using optical and electrophysiological high density recording techniques in awake animals. The role of these features in perception will be identified with behavioural assays. Specific intra- and interareal feedback connections, typically not included in Deep Leaning models, will be opto- and chemogenetically perturbed to assess their contribution to precise nonlinearities of the system and their role in the emergence of complex perceptual structures. Based on these structural, functional and perturbation data, a new generation of well-constrained and predictive sensory processing models will be built, serving as a platform to extract general computational principles missing to link neural activity to perception and to fuel artificial neural networks technologies.

 Publications