Joint laboratory LITIS Saagie – L-Lisa

Saagie's ambition today is to become one of the world leaders in DataOps by offering a solution that orchestrates the best of Big Data and AI technologies to control the data value chain from end to end and accelerate the production of data applications. As an extension of DevOps and DataOps and with the same objectives of collaboration and agility, MLOps now offers practices that take into account the specificities of Machine Learning. In the medium term, Saagie wants to offer a reference platform for MLOps.
The Laboratoire d'Informatique, du Traitement de l'Information et des Systèmes (LITIS) has recognised expertise in machine learning and more particularly in deep learning, namely for sequence and image analysis. This expertise is applied through numerous projects in various application fields: document image analysis and recognition, health, intelligent vehicles, economy, etc. As a result, LITIS needs efficient tools and methods for the rapid deployment of its methodological developments towards industrial or institutional partners.
Saagie and LITIS wish to join their efforts and combine their skills to provide answers to the problems often encountered in the implementation of deep learning techniques, and more particularly on weakly supervised learning and the robustness of models.
One of the main challenges of deep learning lies in the need for large amounts of data to train, validate and operationalize algorithms. Even considering the huge existing data repositories and including the different types of exploitable data (text, images, audio, etc.), the difficulty remains since in most cases, only a small fraction of this data is labeled, and frequently with inaccuracies and errors. If the data from the socio-economic world are growing increasingly, it may be delicate or even impossible to exploit them with statistical methods when the prior annotation work requires too much human effort. Furthermore, although a variety of unsupervised approaches exist, they are often difficult to implement and evaluate. For this reason, state of the art models - whether deep or shallow - are so-called "supervised" approaches requiring labels.
To respond to this problem, weakly supervised learning techniques that reduce the amount of labelled data required are of major interest. In the same time, the available data combine different modalities (images, texts, sounds, symbolic data, etc.) and are characterised by spatial (distance, direction, position) and/or temporal (number of occurrences, changes over time, duration) attributes. These data therefore have multimodal and sequential properties that must often be taken into consideration to meet application needs.
The work carried out in this collaborative laboratory will thus aim at unifying Saagie's technical know-how with the methodological expertise of LITIS so as to develop innovative and efficient algorithms allowing the weakly supervised learning of deep models. The validation will be carried out on two real use cases: i) the automatic diagnosis of lymphoma on data provided by the H. Becquerel center, and ii) decision support based on user data from the Saagie product.