Development of noise-based artifical intellgence approaches

Current approaches to AI are largely based on extensive vector-matrix multiplication. In this postdoctoral project we would like to pose the question, what comes next? Specifically we would like to study whether (stochastic) noise could be the computational primitive that the a new generation of AI is built upon. This question will be answered in two steps. First, we will explore theories regarding the computational role of microscopic and system-level noise in neuroscience as well as how noise is increasingly leveraged in machine leaning and artificial intelligence. We aim to establish concrete links between these two fields and, in particular, we will explore the relationship between noise and uncertainty quantification.
Building on this, the postdoctoral researcher will then develop new models that leverage noise to carry out cognitive tasks, of which uncertainty is an intrinsic component. This will not only serve as an AI approach, but should also serve as a computational tool to study cognition in humans and also as a model for specific brain areas known to participate in different aspects of cognition, from perception to learning to decision making and uncertainty quantification.
Perspectives of the postdoctoral project should inform how future fMRI imaging and invasive and non-invasive electrophysiological recordings may be used to test theories of this model. Additionally, the candidate will be expected to interact with other activates in the CEA related to the development of noise-based analogue AI accelerators.

Co-design strategy (SW/HW) to enable a structured spatio-temporal sparsity for NN inference/learning

The goal of the project is to identify, analyze and evaluate mechanisms for modulating the spatio-temporal sparsity of activation functions in order to minimize the computational load of transformer NN model (learning/inference). A combined approach with extreme quantization will also be considered.
The aim is to jointly refine an innovative strategy to assess the impacts and potential gains of these mechanisms on the model execution under hardware constraints. In particular, this co-design should also enable to qualify and to exploit a bidirectional feedback loop between a targeted neural network and a hardware instantiation to achieve the best tradeoff (compactness/latency).

Application of formal methods for interferences management

Within a multidisciplinary technological research team of experts in SW/HW co-design tools by applying formal methods, you will be involved in a national research project aiming at developing an environment to identify, analyze and reduce the interferences generated by the concurrent execution of applications on a heterogeneous commercial-off-the-shelf (COTS) multi-core hardware platform.

Design of an embedded vision system integrating a fast intelligent imager

The goal of the postdoc is to evaluate the interest of smart imagers integrating processing in the focal plane in embedded vision systems for a localization function and to propose a complete embedded vision system integrating a smart imager and a host.
The study will focus on ego-localization applications, to realize, for example, a 3D localization function.
From an existing application chain, the post-doctoral fellow will be able to carry out an algorithmic study in order to optimize it to exploit the qualities of the intelligent imager.
Then he will be able to propose a partitioning between smart imager and host system, according to performance criteria.
An experiment using the RETINE smart imager as well as the IRIS host board could be conducted to validate the proposal.