Section: New Results

Algorithmic ersatz of cerebral subsystems

As far as the systemic modeling and simulation of high-level brain functions are concerned (e.g., sensory-motor behavior, action selection and planning, perceptual categorization), we need to confront biologically plausible models at different scales of description with functional models that are not constrained by biological facts but still reproduce the expected functional response. This is mandatory to benchmark bio-physical models with respect to their equivalent in classical machine learning, in order to evaluate the degree of naiveness of their performances and also to build feasible simulation in which detailed biological models can interact with less plausible modules in order to be evaluated in realistic numerical situations.

This year, a set of formalism such as the Friston free-energy minimization general principle, deep-learning and related architectures, and more specific formalisms such as harmonic control or adaptive-subspace self-organizing maps have been studied and reviewed. The next step is to write a review, with the challenge of proposing an unifying view of those, and at a more concrete level, to propose the integration of a relevant subset of the related algorithms as a easily usable toolbox. This can be particularly useful to design global models of cognitive functions, even if biologically-inspired models are not yet available for all their components.

Preliminary key points regarding numerical experimentations have been published in this methodological paper [16] .