Neuro-inspired Deep Learning
We use a combination of neuroscience and AI approaches to improve deep learning, and to make it more « human-like ». We design AI systems in which the resemblance with human cognitive functions and brain architecture is explicitly enforced. For example, using NLP data (with T. van de Cruys, IRIT) or brain data (with L. Reddy, CerCo) as a constraint or « regularizer » applied to neural networks throughout their training, we aim to make the neural networks more human-like, and less prone to baffling mistakes (e.g. « adversarial examples » in which a turtle is confused with a rifle). We build novel deep learning models that jointly learn representations across multiple domains and modalities, as humans do: visual recognition and NLP, but also audition, speech recognition, visual reasoning, question answering, action learning and decision-making (Deep RL, robotics); this cross-domain learning, a form of semantic referential grounding, is achieved by adapting existing techniques such as unsupervised neural machine translation or CycleGANs.
In collaboration with the chair of Thomas Serre, we develop new neural network architectures with recurrent and feedback connections, similar to those observed in the brain. For example, « predictive coding » (a neuroscientific theory of feedback brain connections) is adapted to current deep learning frameworks, in order to make the models more robust to difficult situations (e.g. clutter, noise) by iteratively refining the feed-forward inputs through a top-down generative model.
Mathematical analysis (with F. Filbet and G. Faye, IMT) assists us in model optimization, and to investigate spatial and temporal structures (e.g. traveling waves). Math is crucially needed for evaluation and improvement of existing models, derivation of new bio-inspired models and construction of hierarchical links between these models.
Programme : IA collaborative
Thèmes : Apprentissages avec peu de données ou des données complexes, Langage, Neuroscience et IA, Robotique et IA
Rufin Van Rullen, DR CNRS, CerCo
Equipe: Co-Chaire: Leila Reddy (CR CNRS), Tim van de Cruys (CR CNRS), Francis Filbet (Prof. UT3), Gregory Faye (CR CNRS)
PhDs ANITI: Romain Bielawski, Benjamin Devillers, Mathieu Chalvidal (co-direction T. Serre)