Connectionists: PhD offer on time series prediction

[Eric Gaussier]

University Grenoble Alpes and Savoye offer a PhD position on time time
series prediction (see description below), starting ideally in September
2023.

Interested candidates should send a complete CV to Emilie Devijver (
emilie.devijver at univ-grenoble-alpes.fr), Eric Gaussier (
eric.gaussier at imag.fr) and Aurélien Leguy (aurelien.leguy at savoye.com).

The successful candidate will work both at Savoye, in the Advanced Research
and Innovation team, and Univ. Grenoble Alpes, in the Aptikal team. She/he
will also be a member of the Grenoble Interdisciplinary Institute in
Artificial Intelligence (https://miai.univ-grenoble-alpes.fr/).

*******
*PhD description*: Savoye is a fast-growing mid-sized company with around
1,000 employees that offers hardware and software solutions to automatize
warehouses and enhance logistics processes. One of our flagship solution
pertains to high-density robotic storage systems that Savoye pioneered (
https://https://www.savoye.com/us/ressource/ x-pts-shuttles/
<https://www.savoye.com/us/ressource/x-pts-shuttles/>). Savoye is expanding
quickly, thanks in particular to its expansion into the American and Asian
markets (revenue growth >20%/year). Our ambition is to become one of the
main players in the intralogistics sector worldwide in the next few years.  In
2022, Savoye introduced machine learning for the first time in one of its
products as part of a workforce management feature. In order to continue
the efforts undertaken in the field of time series prediction, we propose a
CIFRE whose purpose is to explore and exploit state-of-the-art techniques.
Several topics and strategic issues will support this research :

   -

   —  The prediction of flows and article demands, for example, is at the
   heart of our concerns to optimize inventory management by forecasting
   customers needs,
   -

   —  The estimation of the time required to carry out logistics operations
   (ETA, Estimated Time of Arrival) is a major challenge that can allow for a
   significant improvement in the management of the logistics activities in
   the warehouse (taking into account the effect of routing decisions on ETA),
   -

   —  Forecasting of failures and break-down of equipment is also a
   critical topic with a view to expanding our maintenance service offering.

To address these issues, we will draw on recent developments in artificial
intelligence on deep neural networks and causality. In particular, we want
to explore transformer-based models ([5]) for flow and demand prediction
and time estimation, and discovery and causal inference models for fault
prediction ([1]).  One of the major problems lies in the fact that these
models have excellent performance on data close to those used for their
training and often poor performance on data that differs from those used in
training. We intend to solve this problem by using generic causal
representations ([6]) because they are representative of the underlying
physical phenomena. The causal graph and associated latent variables can be
partly extracted from the available data from discovery and causal
inference methods ([1]). This will improve the genericity of models and
their transferability to different contexts. In addition, the available
data are often partial, with missing values, and are based on different
trends that are repeated at more or less regular intervals. We introduced
the notion of pseudo-periods to designate the fact that repetitions are not
entirely regular and proposed a mechanism capable of identifying them with
recurrent neural networks as LSTMs ([2]). This thesis will also be an
opportunity to explore, on the one hand, new methods for the identification
of trends and pseudo-periods to improve predictions and, on the other hand,
new methods to optimize the use of partial or missing data (e.g. data
produced during lockdowns).

It should be understood that these topics/themes are neither fixed nor
exhaustive but constitute a basis for discussion and work.
*Savoye Advanced Research and Innovation Service*: The PhD student will be
part of the Advanced Research and Innovation team, which reports to the
group’s top executives. Its scope is to explore the state of the art of
research to identify and explore (by prototyping or carrying out proofs of
concept) solutions that may ultimately significantly improve the company’s
products or propose radically new solutions. She/he will work within a team
specialized in AI in the broad sense (from operations research and
mathematical optimization to different ML techniques) in direct
collaboration with the AI/ML tech lead and will aim to extend the already
in-depth research that has been conducted on the subject. Amazon’s research
illustrates (but does not guide) our trajectory from a technological
perspective ([4]). Experiments based on models focused on attention
mechanisms (transformers) are at the heart of our current research
activities. We are also interested in exploring innovative approaches like
DeepMind for Google Maps ([3]).

*University Grenoble Alpes*: The person recruited will be integrated into
the APTIKAL team of the Grenoble Computer Science Laboratory and will be
co-supervised by Émilie Devijver (CNRS researcher, mathematician) and
Éric Gaussier (professor of computer science at the University of Grenoble
Alpes, Director of MIAI - Multidisciplinary Institute in Artificial
Intelligence). The APTIKAL team is specialized in machine learning and
knowledge acquisition and in particular on statistical learning and
causality, with both broad applications and more targeted applications on
natural language processing or information retrieval for example.
*******

-- 
Eric Gaussier
Grenoble University, LIG Laboratory
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20230524/088c992e/attachment.html>