> Academic opportunities > PHD positions

Low-power AI for uncertainty modelling

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

Department: Département Systèmes et Circuits Intégrés Numériques (LIST)

Laboratory: Laboratoire Intelligence Artificielle Embarquée

Start Date: 01-09-2022

Location: Saclay

CEA Code: SL-DRT-22-0355

Contact: thomas.dalgaty@cea.fr

Models of AI, notably deep neural networks, are increasingly incorporated into practical systems in a diverse set of domains. In some of these domains, autonomous vehicles and wearable medical devices for example, these models are responsible for taking safety-critical decisions. However, the models that are prevalent today have a big problem: they are not able to reliably communicate or understand when they are uncertain about a prediction. It is becoming increasingly clear that, for these models to be applied responsibly in certain applications, it is vital that their uncertainty is taken into account in the decision making process. Various approaches have been developed that allow neural networks to model their uncertainty (i.e., Bayesian and ensemble methods). However, each of these entails significant increases in the memory, energy and time required to run and train these models making their use impractical in the context of edge and in-sensor AI systems that are often heavily constrained in memory and energy. This is in part due to a lack of consideration in how hardware solutions might optimised to run uncertainty models ? something that this PhD subject will seek to remedy. The student will work on the problem of applying uncertainty models to a new dataset recently collected by CEA-LETI Clinatec (http://www.clinatec.fr/en/) composed of electroencephalography time-series recordings from patients suffering from epilepsy. Specifically, the student will: develop a benchmark task using this dataset, evaluate different approaches on this benchmark, propose and study the impact of complexity reduction techniques and most importantly propose hardware innovations to improving the energy and memory efficiency of the most promising approaches.

See all positions Download the offer (.zip)

Email Bookmark and Share