EN FR
Homepage Inria website

New Software and Platforms
Bilateral Contracts and Grants with Industry
Bibliography


Section: Research Program

Interpretability

The pervasiveness of complex decision support systems, as well as the general consensus about the societal importance of understanding the rationale embedded in such systems (General Data Protection Regulation, recital 71 http://www.privacy-regulation.eu/en/r71.htm), has given momentum to the field of interpretable ML. Being a team specialized in data science, we are fully aware that many problems can be solved by means of complex and accurate ML models. Alas, this accuracy sometimes comes at the expense of interpretability, which can be a major requirement in some contexts (e.g., regression using expertise/rule mining). For this reason, one of the interests of LACODAM is the study of the interpretability-accuracy trade-off. Our studies may be able to answer questions such as “how much accuracy can a model lose (or perhaps gain) by becoming more interpretable?”. Such a goal requires us to define interpretability in a more principled way —an endeavour that has been very recently addressed, still not solved. LACODAM is interested in the two main currents of research in interpretability, namely the development of natively interpretable methods, as well as the construction of interpretable mediators between users and black-box models, known as post-hoc interpretability.

We highlight the link between interpretability and LACODAM's axes of decision support, and user/system interaction. In particular, interpretability is a prerequisite for proper user/system interaction and is a central incentive for the advent of data visualization techniques for ML models. This convergence has motivated our interest in user-oriented post-hoc interpretability, a sub-field of interpretable ML that adds the user into the formula when generating proper explanations of black-box ML algorithms. This rationale is supported by existing work  [28] that suggests that interpretability possesses a subjective component known as plausibility. Moreover, our user-oriented vision meets with the notion of semantic interpretability, where an explanation may resort to high level semantic elements (objects in image classification, or verbal phases in natural language processing) instead of surrogate still-machine-friendly features (such as super-pixels). LACODAM will tackle all these unaddressed aspects of interpretable ML with other Inria teams through the IPL HyAIAI.