Biologically Inspired Machine Learning

Nowadays, the field of machine learning is experiencing exponential growth and interest. The recent advances in deep learning have changed the everyday life of many users of smartphones, for instance, and will also change everyone’s life in a near future with autonomous cars. These algorithms are actually inspired by neuronal networks and have undergone several improvements (algorithm stability and reduced energy consumption by hardware devices, for example) that are mainly due to the injection of biological concepts such as neuronal inhibition inspiring dropout regularization strategies or temporal synaptic integration.

On the other side, there is still no complete mathematical understanding of why such algorithms are reliable. Several promising results have already been obtained, in particular the improvement of such algorithms through active learning strategies to significantly reduce the amount of training data required and through bio-inspired network pre-training techniques. Thanks to a constant dialog between computer scientists, mathematicians, biologists and psychologists, researchers from Université Côte d’Azur aim to:
  • inject more biologically inspired features in these algorithms such as synaptic plasticity or neuronal assembly generation;
  • evaluate the performance of an algorithm on more human/cognitive aspects;
  • improve the understanding of such algorithms thanks to the local mathematical expertise in mathematical neuronal models but also in pure mathematics such as differential geometry, logic or algebraic semantics;
  • find new self-learning software/hardware implementations.