Since 10 yers, deep learning improves the state of the art in multiple domains as image classification, natural language processing, game playing, ... However, as these models perform statistical learning from datasets, supposed representatives of the task, they suffer from some limitations as a lack of generalization to different contexts and a lack of robustness, as e.g. minor modifications in input pictures can largely decrease the classification performances. In practice, these problems may limit the use of these technologies in real world applications, or even pose major risks (e.g. autonomous cars) and thus affect the trust we have in these systems. Theoretically speaking, this tends to show that the patterns learned by these models do not truly correspond to the concept of objects.

We propose to tackle these limitations by taking inspiration of the cognitive processes in animals, including humans. The main idea emphasized by the studies in cognitive science, neuroscience, philosophy, ... is that cognition is embodied and that perception is not a fixed process of gathering statistical correlations but rather an active and exploratory one. Thus, there is a fundamental misalignment between human and typical AI representations: while the former are grounded in rich sensorimotor experience, the latter are typically passive and limited to a few modalities.

Especially, we are interested in the sensori-motor contingencies theory that combines these various aspects in an unified framework. The two main claims are the sensori-motor contingencies are “the structure of the rules governing the sensory changes produced by various motor actions” and that “perceiving [...] is an organism’s exploration of the environment that is mediated by knowledge of SMCs”.

In this thesis, we propose the learning, by an autonomous (i.e. unsupervised, generic and adaptable) agent, of efficient and resilient representations by applying the core concepts of the sensori-motor contingencies theory to deep learning methods in order to overcome their current fundamental limitations.