Title: HPC for learning in tree-based tensor format
Date: 2022-04-20 09:21
Slug: job_ca09cf78a33f378a4ae56d3fd701fbfa
Category: job
Authors: Paul Mycek
Email: mycek@cerfacs.fr
Job_Type: Post-doctorat
Tags: postdoc
Template: job_offer
Job_Location: Paris (Issy-les-Moulineaux)
Job_Duration: 24 months
Job_Website:
Job_Employer: Airbus-Cerfacs-Inria
Expiration_Date: 2022-07-31
Attachment: job_ca09cf78a33f378a4ae56d3fd701fbfa_attachment.pdf
In the context of a joint team between Airbus C R & T, Cerfacs and Inria a 24-month postdoc position is available in Paris (France).
In the context of supervised learning (e.g., learning regression models as approximation of functions) or unsupervised learning (probability distributions learning), we face CPU time issues as the dimension grows.
We encounter these problems of supervised learning in high dimension when we are interested in the prediction of physical quantities, which are very often spatial fields. As an example, we can mention the learning of wall laws in a fluid calculation, which allows us not to refine too much near a wall.
The unsupervised context is frequently encountered in the industrial context, for example when we want to predict a flight quality score from data coming from sensors of the aircraft, and when we want to find an inconsistent or abnormal behavior of the aircraft. We then try to find anomalies in a data set and the dimension can once again be very large.
The tools we are investigating and developing are known as the Continuous Graphical Model (https://hal.archives-ouvertes.fr/hal-03346955/document) and the Deep Tensor Network (https://arxiv.org/pdf/1811.04455.pdf). For both Continuous Graphical Models and Deep Tensor Trees, some operations are particularly costly:
- Algebraic operations on tensors.
- The exploration and quantization of many tree or graphical model configurations (Directed Acyclic Graphs).
- In the case of tensor networks, learning the coefficients in the leaves of the tree is also expensive (alternate least-squares: https://arxiv.org/pdf/1305.0030.pdf).
The objective of this post-doc is to accelerate these three functionalities by using different strategies.
This position is intended for candidates with a strong background in computational sciences, preferably holding a PhD in applied mathematics, with some knowledge/experience in machine learning, probability, numerical linear algebra.
For more information, please refer to
https://recrutement.inria.fr/public/classic/en/offres/2022-04835
or contact
Pierre Benjamin (Airbus CRT, pierre.benjamin@airbus.com)
Olivier Coulaud (Inria, olivier.coulaud@inria.fr),
Paul Mycek (Cerfacs, mycek@cerfacs.fr).
This 24 month position is planned to start on October 1st, 2022 at the very latest.