A growing number of embedded applications, confronted with varying and non-controlled environments, require an increased degree of adaptability and analysis capabilities to fulfill their task. Preprogrammed actions are no longer able to deal with these new sets of tasks, and are therefore being replaced by a promising paradigm: deep learning.

However, deep neural networks are susceptible to data distribution shifts occurring between training and use. This apparent flaw prevents the widespread deployment of deep networks in embedded products. Furthermore, it is impossible to gather and add enough data to the training set to cover for all possible shifts, due to their tremendous diversity.

The origin of this vulnerability lies in part in the shortcut-learning behaviour of deep networks: they learn only the most efficient patterns, no matter how spurious, and completely disregard the others. Confronted with a new distribution, in which the predictive patterns are partially different, a network that learned a limited subset of features will be less likely to be able to make a proper decision.

In collaboration with Thales Land and Air Systems, the aim of this work is therefore to develop solutions for mitigating the domain shifts performance drop in deep networks. This work has two main contributions.

Firstly, we propose a new deep generative architecture that mitigates the shortcut-learning behavior in an under-explored setting. Previous state-of-the-art works relied explicitly on shortcut-contrary samples and increased their importance in the training procedure. In this work, we demonstrated on several different synthetic benchmarks that such particular samples were not needed for shortcut avoidance, and further confirmed the effectiveness of our approach on a realistic benchmark.

Secondly, the work presented here deals with a more general domain shifts situation, in which only a single domain is available during training. Test-time adaptation has emerged as a promising set of strategies to efficiently increase performance when facing new domains at use time. They however rely on a model trained with the standard procedure, which, as previously stated, ignore some predictive patterns. We propose a training-time approach complementary with test-time adaptation. Our method seeks to learn both the patterns learned through standard training and the normally "hidden" ones, and as a result enable a more thorough test-time adaptation. Based on extensive experiments, we show that our approach improved the quality of predictions on domains unseen at training-time.