UNSUPERVISED LEARNING OF PATTERNS IN THE SENSORY CORTICES: PROBBING FUNCTIONAL CONNECTIVITY MODELS AND HEBBIAN LEARNING DYNAMICS – ULEARNINGCTX

Most of the life time of an animal consist in exploring and sensing its environment without any specific goal. During this process, the animals encode the statistics of the environmental stimuli that surround them and adapt progressively their neuron responses to these stimuli. This unsupervised learning of the environment is thought to happen in the cerebellar cortex: sensory cortices encode the stimuli using Hebbian learning rule, i.e. assuming that the coactivation of a pattern in a network produces a change in the synaptic strengths between the involved neurons. Even if it exists several signatures of Hebbian learning, the validity of these leaning rules is still a matter of debate and several other learning rules have been proposed. How do the sensory cortices encode memories, patterns or stimulus associations is still an open question.
With the recent development of new optical tools to record (multiphoton microscopy) and manipulate (optogenetics) the brain activity, it is now possible to imprint patterns of neural activity in the cortex of mice eliciting behavior. In parallel during the last decade, unsupervised machine learning tools have been developed in order to infer the effective connectivity matrix of a network from its activity, or even, to fit on neural activity, Hopfield models based on Hebbian learning using the restricted Boltzmann machine inference.
In this project, we propose to use our ultra-fast two-photon microscopy, optogenetics perturbation and modelling techniques to understand the unsupervised learning rules and dynamics in cortical networks of mice. A series of neuronal ensembles are activated optogenetically until their fixation in the network. How fast does the cortex imprint a series of pattern and build attractor states? What is the exact underlying Hebbian learning rule? What is the validity of effective connectivity models?