Abstract :
[en] The present thesis takes a step towards enriching artificial neural networks with bio-inspired mechanisms. To this end, high-level abstractions of important biological rules, modelled using control theory, will be introduced and linked to artificial networks. In particular, it will be discussed that introducing neuronal bistability and neuromodulation into artificial neural networks provides different benefits. As a first step, Part 1 will first introduce necessary machine learning background.
Part 2 of this thesis will focus on the ability of recurrent neural networks to learn long-term dependencies, something which usually proves difficult. Bistable recurrent cells will be introduced as a way to help towards solving such issues. Furthermore, supported by the results obtained with those cells, a more generic method to promote multistability with usual recurrent cells is proposed. This part highlights the importance of dynamics in recurrent neural networks and in particular, right after initialisation, for easily learning long-term dependencies.
Part 3 of this thesis is dedicated to introducing neuromodulation in artificial neural network. This important biological mechanism is often associated to the robust control of continuous behaviours, allowing biological systems to adapt very quickly to changing context, something which remains very difficult for usual artificial agents. As such, a neuromodulated architecture, specifically designed for its adaptive capabilities (i.e. robustness towards changing environment or context), is proposed. It will be shown to exhibit much more stable performance and to converge towards better policies than classical recurrent networks. Furthermore, this part discusses that these architectures are also implicitly able to learn a continuous representation of the different contexts in which they evolve. Finally, some other very recently proposed architectures and their benefits will be briefly mentioned as well.
As a last note, it is important to mention that Part 2 and Part 3 of this thesis could potentially be linked. As such, the last part of this thesis will be dedicated to provide ideas for future works, specifically aimed at closing the gap between Part 2 and Part 3.
