Better Representations for Artificial Intelligence (BRAIn) is a research team hosted at IMT Atlantique and part of the Mathematical and Electrical Engineering Department. It is also part of the CNRS laboratory Lab-STICC.
The purpose of BRAIn is to investigate key questions at the crossbreed of Artificial Intelligence, Deep Learning and Signal Processing, with applications using images, sounds, text and more complex domains including neuroimaging data.
As of 2022, these questions include:
- Few-shot learning
- BRAIn is ranked #1 on miniImageNet (c.f. PEME-BMS method: link)
- Continual (or incremental) learning
- BRAIn won the NIC (New Instances and Classes) continual learning competition at CVPR 2020
- Predicting generalization in deep learning
- BRAIn ranked #3 at the PGDL competition at NeurIPS 2020 (link)
- Compression of Deep Neural Networks
- BRAIn collaborates on this subject with Mila (lab directed by Yoshua Bengio) in Montréal, Polytechnique Montréal
- Some specific targets: autonomous vehicles, drones, microcontrollers, smartphones
- A particularly visible post of the team on towardsdatascience: https://towardsdatascience.com/neural-network-pruning-101-af816aaea61 with 300+ claps
- Graph Neural Networks and Graph Signal Processing
- Transfer Learning and Self-supervised Learning
- Application to brain encoding models of auditory cognition – with the Courtois Neuromod project (Pierre Bellec, Université de Montréal)
- Representation learning of audio signals
- Associative memories
- BRAIn collaborates with University of Münster, University of Brest
- Robustness and Neural Network optimization
- BRAIn collaborates with University of Southern California
- Artificial Intelligence and Ethics
BRAIn was created in the continuation of the NEUCOD project (led by Prof. Emeritus Claude Berrou) funded by the European Research Council (ERC FP7 290901), as well as the Neural Coding and Neural Communications project funded by the Britanny Region and the CominLabs LabEx.