Context
This post doc takes place in the DATAWISE (Data Annotation Technology Advancement With Innova-
tive Solutions for Efficiency) project between Neovision, a company in Grenoble, and the LIRIS laboratory
in Lyon. It targets the use of self-supervised learning methods to obtain transferable image representations. In particular, we expect a structured space, that can be used in an assisted labeling software
developped by Neovision, even when applied to industrial data of Neovision’s partners, that can be
largely different from images of classical datasets.
Subject
Modern self-supervised learning (SSL) methods consist in defining multiple augmented versions of
inputs images and forcing their representations (i.e. the embedding of a deep visual architecture) to
be similar. The underlying idea is that the augmentations modify some factors of the images (e.g.
color, position, …) that are not related to the content of the image (that we want the model to
learn). Thus, it defines a space of representations that is invariant to these augmentations, i.e.
related to the semantic of the image. In we introduced the parameters of the transformations in
the representation, via an equivariance module, to improve visual classification of SoTA architectures,
especially in transfer.
However, preliminary experiments performed in the project tend to show that the transfer is not
really efficient toward industrial datasets. This may be due to use of the same augmentations strategy
in all methods, that can be somehow overfitted for the classical datasets. Some work try to generate
dedicated augmentations e.g., and we want to explore if such strategy can improve transferability.
Thus the work will consist in:
- Extend automatic augmentation generation strategy, including ones proposed by a previous post
doc, and evaluate them on transfer to industrial data
- Study how these augmentations can leverage representations whose features can be somehow
prompted/filtered with expert knowledge on the data, for instance to perform an additional ne
tuning on the industrial dataset
Profile
Ideally, the candidate would have the following skills:
– background in artificial intelligence / machine learning / computer vision (or equivalent)
– previous experience in deep learning especially in self-supervised models and/or application to
image data
– good programming skills (especially in Python and Pytorch/Tensorflow)
– interested in applied research
Localisation
Doua campus, LIRIS laboratory, Lyon, France. The post doc will also interact with employees of
Neovision, Grenoble, France.
Duration
9 months (with a potential follow up on other projects), starting as soon as possible.
Salary
Around 2300 €/month net salary.
Advisors
Stefan Duffner: professor, LIRIS laboratory, Imagine group, Lyon
Mathieu Lefort: associate professor, IRISA laboratory, MaLT group, Rennes
Application
To apply, please send a CV, an application letter and recommandation letters to Stefan Duffner (stefan.duffner@liris.cnrs.fr). Applications will be studied when received.
If you have any question regarding this position, please send an email to Stefan Duffner.
