Building more accurate AI: ReciTAL and ESILV launch the Research Chair in Bayesian Neural Networks

Thanks to this new Research Chair in Bayesian Neural Networks, ReciTAL, the French AI startup specialised in Document Intelligence solutions, and ESILV are joining forces and expand their expertise in the field of Artificial Intelligence.

One of the criticisms about AI is the incapacity of current ML models to represent the concept of uncertainty accurately. In other words, they are unable to make reliable predictions. 

Bayesian deep learning allows neural networks to improve the degree of reliability of their predictions by adding uncertainty during the training phase and the prediction phase of the system.

ReciTAL seeks to speed up its research on VQA (Visual question answering) models, to allow more effective human-machine interaction. Therefore, the AI-based platform has approached ESILV, and more specifically, the De Vinci Innovation Center, to create the ReciTAL Research Chair in Bayesian neural networks.

Bayesian model, a method to avoid the “black box” effect of deep learning

While artificial intelligence has made more and more progress in many areas, it keeps raising questions and doubts, particularly when it comes to safety-critical applications such as the autonomous car or health.

The most recurring criticism of this technology is the failure of humans to fully grasp the working model, to explain it in simple language and to confront it with a human decision-making process.

To explain this lack of interpretability of the algorithms, some people call it a “black box” effect. We know the data that goes into it and what results come out of it, but we do not know what is going on within it. The Bayesian Machine Learning or Deep Learning is aimed at addressing this need for clarity and at overcoming the inability of existing ML models to represent the concept of uncertainty reliably.

“We are trying to make Artificial Intelligence more “intelligent”. It is a cognitive function that is indispensable for its use in everyday life to overcome its black box behaviour but also to facilitate any interaction with a human being”. (Clément Duhart, research professor at ESILV, in charge of the DVIC)

The Bayesian paradigm is based on an updating of the probability distribution on the models after observing a training data set. Therefore, it would enable ML models to represent the notion of uncertainty reliably. Although the Bayesian model has gained interest among researchers, concrete applications are still to come.

ReciTAL and ESILV, a win-win partnership

ReciTAL is an IA startup specialised in Document Intelligence and Automatic Language Processing ( TAL) solutions, which has been launched in 2017. The French startup was listed in the Teknowlogy Group’s global report “AI Vendors of Tomorrow”.

ReciTAL’s research work mainly focuses on : 

• Document Layout Understanding: it uses models which combine textual and visual features to automatically detect the internal layout of documents and facilitate document analysis and use.
• Small data and active learning: it reduces the data to improve the performance of models.
• Question Answering and Question Generation: The full range of approaches involving the transition from keyword questions to natural language questions.

VQA (Visual Question Answering) can help artificial intelligence in terms of visual perception and natural language communication. It is a visual question-answering system that will transform the way machines understand the content of images. It is in order to provide more insight into the recent problem of VQA that ReciTAL approached ESILV as part of this R&D partnership.

The Chair will host a PhD student, co-funded by ReciTAL, who will work for 3 years on the Bayesian approach to machine learning with a specific application on VQA systems. The budget allotted to the research equates to €150k over 3 years.

Aware of the growing importance of artificial intelligence, ESILV is now proposing the Data and Artificial Intelligence specialisation as of this year. Engineers trained in this program are experts in the use of data, whether big or small, from its collection, modelling and storage to its analysis and interpretation. The courses in this specialisation revolve around Machine Learning and Deep Learning technologies, as well as statistics and IT, IT project management, cloud computing and agility.

Learn more about ESILV.