Organisation/Company Université Côte d'Azur Department Mathematics Research Field Mathematics » Applied mathematics Computer science » Other Researcher Profile Recognised Researcher (R2) Positions Postdoc Positions Country France Application Deadline 31 Dec 2024 - 11:55 (Europe/Paris) Type of Contract Temporary Job Status Full-time Offer Starting Date 1 Jan 2025 Is the job funded through the EU Research Framework Programme? Other EU programme Is the Job related to staff position within a Research Infrastructure? No
Offer Description Context: In recent years, machine learning algorithms have become capable of competing with and even surpassing humans in many tasks related to artificial intelligence. Problems long considered beyond the reach of computers are now routinely solved by algorithms built on the precepts of machine learning: collect large quantities of data and train on them. The drawback of this approach is the increasing complexity of these models. Perhaps the simplest way to perceive this growth is to look at the progression in the number of parameters in these models: in recent years, it has become commonplace to train models that depend on billions, or even hundreds of billions of parameters. But the size of these models doesn't seem to be the only driver of their improved performance, and their architecture has also become incredibly complex. Machine learning has moved on from simple geometric and statistical ideas to complex non-linear architectures, particularly since the renaissance of deep neural networks. These two factors have led the community to refer to these models as "black boxes", which is sometimes reinforced by the fact that the implementation of these models is not publicly accessible. All these factors contribute to making it very difficult to understand how a particular prediction is made by such a model. The ANR JCJC NIM-ML is dedicated to the development of new-generation interpretability methods.
Mission: The researcher recruited will be in charge of developing the 'time series' activity of the project, i.e. studying and developing interpretability methods for algorithms taking time series as input. The aim of studying existing methodologies is to prove that they make sense in simple cases, or on the contrary, to show where they fail. The researcher recruited will continue the project's recent work in this direction. The development of new methodologies will respect this imperative of good theoretical guarantees on simple models. Proven statistical methods will be used to segment the underlying signals. Beyond the post-hoc point of view (already trained), which is beginning to be well understood, we'll be moving towards approaches that can be interpreted 'by nature', such as transformers.
Activities:
state of the art on current approaches to time series interpretability
extension of recent work carried out within the project to 'post-hoc' methods such as Anchors for transformers
development of new methods based on break detection
extension to the 1D domain (not necessarily temporal, e.g. electrocardiograms)
Skills:
excellent writing skills
perfect command of English
in-depth knowledge of Python and the PyTorch framework
Diploma required: PhD in mathematics, computer science or similar fields
Contract duration: 1 year
Desired start date: January 1, 2025
This job is part of the NIM-ML project (ANR JCJC)
#J-18808-Ljbffr