> Academic opportunities > PHD positions

Simplifying Artificial Intelligence: learning interpretable representation

Technological challenge: Artificial intelligence & Data intelligence (learn more)

Department: Département Systèmes et Circuits Intégrés Numériques (LIST)

Laboratory: Laboratoire Intelligence Artificielle Embarquée

Start Date: 01-10-2022

Location: Saclay

CEA Code: SL-DRT-22-0305

Contact: laurent.soulier@cea.fr

The artificial intelligence field is dominated today, in volume, by research on deep neural networks (DNN). Indeed, about a decade ago, they seemed to significantly improve on problems deemed as difficult, such as computer vision issues. Since then, DNN are generating enthusiasm and hopes. Yet, even if the trend continues to this day, one must admit that they pervade from laboratory to application in only a few cases. Multiple reasons are explaining this fact. One of those is that DNN technics are computationally and energetically costly (hundreds of watts vs about 15 W for the human brain). Besides, DNN are ?black box? algorithms: when they (don't) work, it is virtually impossible to explain their outcome, which is an hindrance for their design and, especially, for their adoption within critical systems. Another way exists, known as symbolic machine learning. It consist mainly on rules-based system, whose rules are written by human experts. They can be easier to interpret but they generally cannot scale to ?large size? problem. In order to address some of above criticism, we are aiming at taking advantage of the best of both worlds, by designing a method based on state of the art machine learning, not to solve a given problem, but to generate a symbolic algorithm, in the form of an automata, whose function will be to address the problem in an interpretable way, and with a low footprint. A first preliminary way to proceed would be to ?dissect? a deep network, trained for a given task, to extract the most important features for the optimised cost function computation (some layers of a convolutional network) and to isolate the relationship between these features (how the classifying layers, a fully-connected network for instance, is combining the features to produce an output). This work will be valorised into an autonomous robotic application and/or a vision system. The student is expected to have a good knowledge of machine learning and a taste for research within a pluridisciplinary context.

See all positions Download the offer (.zip)

Email Bookmark and Share