Télio Dupuis
| Position: | PhD student |
|---|---|
| Period: | from 2025 |
| Address: | Télio Dupuis |
| LJK - Bâtiment IMAG | |
| 150 place du Torrent | |
| Campus St Martin d'Hères | |
| 38400 Grenoble France | |
| Email: | telio.dupuis@univ-grenoble-alpes.fr |
| Supervisors: | Jean-Charles Quinton, Mathieu Lefort, Frédéric Armetta (LIRIS, Lyon) |
Research description
The proposal deals with the learning of representations from sequences of interaction with the environment. In particular, we will draw on sensorimotor contingency theory [3,4] to ensure that action structures both the representations learned and the dynamics of interaction. Within this framework, we aim to learn predictive structures of the world, enabling self-supervised definition of objects as graphs of potential interactions [5]. During the PhD thesis, the following issues will be addressed:
- How to integrate action into existing self-supervised deep learning models (e.g. Transformer or State Space Models) and what is its influence on the structures and predictive capabilities of the model.
- How to learn spatio-temporal structures that may correspond to notions of proto-objects. Hybrid approaches combining graphs and deep learning will be studied, in particular to learn multiscale, locally organized and globally connected structures. These representations could also serve as a supervisory signal for the self-supervised approaches used for multimodal learning in another part of the project.
- How to obtain efficient methods in terms of learning time and data used. In practice, the use of action requires a simulator which induces longer computation times than the use of datasets. The possibility of offline pre-learning (e.g. with pre-recorded random behaviors) will be investigated. In addition, active learning mechanisms (by choosing the right action to obtain useful information) will be proposed to reduce the amount of training data required for a given performance. By formalizing testable hypotheses about the environment, these mechanisms will also help to reduce the size of representations (by retaining only predictable sub-parts of the inputs). This research could also be coupled with policy choice mechanisms explored in another part of the project.
These different directions will be tested in simple environments (as we have done with Tetris [6]), or in a robotic simulation environment with objects of simple shapes and properties (in line with other research carried out in the MeSMRise project).
References
- Nicholas Hay, Michael Stark, Alexander Schlegel, Carter Wendelken, Dennis Park, Eric Purdy, Tom Silver, D Scott Phoenix, and Dileep George. Behavior is everything: Towards representing concepts with sensorimotor contingencies. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.
- Linda Smith and Michael Gasser. The development of embodied cognition: Six lessons from babies. Artificial life, 11(1-2):13-29, 2005.
- J Kevin O'regan and Alva Noë. A sensorimotor account of vision and visual consciousness. Behavioral and brain sciences, 24(5):939-973, 2001.
- Erik Myin and J Kevin O'Regan. Perceptual consciousness, access to modality and skill theories. A way to naturalize phenomenology? Journal of consciousness studies, 9(1):27-46, 2002.
- Jedediah WP Allen, Bartuğ Çelik, Mark H Bickhard. Age 4 transitions: Reflection as a domain-general development for explicit reasoning. Cognitive Development, 2021.
- Jean-Charles Quinton. Emergence of space from sensorimotor invariants: anticipatory network analysis in the context of the Tetris game. Cognitive Processing, 13, 48, 2012.