Description of the thesis topic

In the field of neuromorphic artificial intelligence, Energy-Based Models (EBMs) are experiencing growing interest due to their ability to capture relationships between variables by minimizing a single scalar function (energy function) without requiring normalization at the scale of the entire dataset. A prominent example of an EBM is certainly the Hopfield network model where the evolution of the state of each unit (neuron) is dictated by minimizing its contribution to the total energy of the system. Once equilibrium is reached (the energy minimum), the result readout can be performed.

EBMs exhibit characteristics that make them attractive for "near-physics" implementations (analog/mixed signal). Indeed, any dynamical system possessing an evolution law whose parameters can be adjusted can be used to create an EBM model performing an arbitrary task, such as image classification. An analog electrical system based on non-volatile memories is well-suited for such tasks, and some initial demonstrations show the interest of this approach for energy efficiency.

Combining these EBMs with more biologically plausible and/or locally-based learning algorithms constitutes a very promising research direction for energy-efficient and Edge AI applications: the goal here is to make these systems capable of learning per se, not just inference.

This doctoral thesis topic builds on work conducted within the laboratory, on a software framework [2] as well as contributions to algorithmics and hardware architecture [3][4][5] that are about to be the subject of a demonstrator circuit realization. This doctoral thesis will address both the question of models and training algorithms and will analyze the opportunities for analog/mixed material implementations and memristive component-based (OxRAM, FeRAM, FemFET) for storing the weights of these networks. These investigations will be conducted within the framework of the "Emergences" project [1] of the national PEPR IA program of France 2030. These investigations will be carried out in collaboration with other partners, doctoral students, and post-doctoral researchers.

Candidates must have skills in at least two of the following areas:

Machine learning and associated mathematical foundations
Embedded systems
Analog/mixed design
[1] https://emergences.pepr-ia.fr
[2] https://www.frontiersin.org/articles/10.3389/fncom.2023.1114651/full
[3] https://hal.science/lirmm-04959185/
[4] https://hal.science/lirmm-04959178/
[5] https://hal.science/hal-05098393/

Work Context

The thesis will be conducted in close collaboration with other partners of the targeted project supported by the PEPR IA (Emergences project). An international collaboration with the University of Bremen is also planned, a university with which collaborative work on EBMs has been conducted for several years.

The position is located in a sector under the protection of scientific and technical potential (PPST), and therefore requires, in accordance with the regulations, that your arrival is authorized by the competent authority of the MESR.

Constraints and risks

NA