A unique scientific partnership centred around AI
What ways are there of building trust among users of artificial intelligence (AI) systems? This is the major question occupying those members of Europe’s scientific community involved in TAILOR: Foundations of Trustworthy AI -Integrating Reasoning, Learning and Optimization. “The aim is to enable the development of new ideas and solutions in the field of trustworthy AI by encouraging researchers to meet up and communicate with each other in order to form unique partnerships”, explains Marc Schœnauer, director of research in charge of AI at Inria and head of the TAU project team. A whole host of conferences, workshops and summer schools have been organised through the network, in addition to special challenges and hackathons. Launched in September 2020 and coordinated by Linköping University in Sweden, TAILOR brings together specialists in neuro-symbolic AI, trustworthy AI and multi-agent systems, industrial partners, and four Inria project teams: LACODAM, MULTISPEECH, ORPAILLEUR and TAU. More than 55 European partners are involved in the project, from 21 different countries. “Cooperating at a European level allows new solutions to be found, while also creating the right conditions for young prospects, helping to tackle brain drain.”
TAU - a project team committed to research at a European level
TAILOR is not the only EU project which TAU is involved in. The team is also involved in TRUST-AI, for example. This research project on the subject of trustworthy AI relates to one specific technique used in machine learning, genetic programming, which can be used “to track all of the choices made by an algorithm” in order to obtain intelligible and explainable models. More specifically, TAU is concerned with predicting energy consumption. Another EU initiative that TAU is a partner in is ADRA-e, a Coordinated support action (CSA) created to support the public-private partnership AI, Data and Robotics. The role of ADRA-e is to assist the private partners involved in the project and to organise meet-ups, workshops and conferences for them to attend. Marc Schœnauer will be in charge of the scientific component of this initiative, which is set to get underway in July 2022.
Developing trustworthy AI systems
The main topic of research uniting the partners involved in TAILOR relates to the development of trustworthy AI. AI must meet a range of criteria - all of which are areas of focus for researchers - in order to be considered trustworthy: it must be robust, transparent, explainable, fair and non-discriminatory, respecting privacy while acting responsibly and having a positive impact on society.
If systems lack transparency, are biased, are easily affected by noise or attacks, or if their results cannot be explained, then society will never accept AI.
TAU project-team leader
If smart machines are to play a role in our lives, then the development of trustworthy AI is vital. “If systems lack transparency, are biased, are easily affected by noise or attacks, or if their results cannot be explained, then society will never accept AI”, stresses Marc Schœnauer. “The same is true for the economy: in industry, no one is going to launch headlong into technology they don't trust. But technology transfer is vitally important if we want to prevent ourselves becoming completely dependent on the USA or China. We’re too late when it comes to the major research platforms, but we can still have our say in sectors such as health or logistics.”
Learning or reasoning ?
“Everyone is working on trustworthy AI nowadays”, continues the researcher. “What sets TAILOR apart is that it is focused on the so-called ‘hybrid’ approach, using optimisation to combine a symbolic approach (reasoning) with a sub-symbolic or digital approach (learning).”
Digital AI reacts better to ‘noise’ and uncertainty, but it still suffers from a lack of explainability and certification: this is an opening where symbolic AI can step in.
Responsable de l'équipe(projet TAU
Historically, these two schools have long been in opposition to each other. Popular in the 1960s, symbolic AI involves simulating human reasoning via a sequence of logical operations. Based on the principle of machine learning, digital AI really began to take off in the 2010s as volumes of data and the power of machines increased, enabling the emergence of Deep Learning. “Digital AI reacts better to ‘noise’ and uncertainty, but it still suffers from a lack of explainability and certification: this is an opening where symbolic AI can step in”, explains Marc Schœnauer.
Inria in charge of the roadmap and challenges
Seven specialists in natural language processing (NLP), time-series prediction, causal modelling and neural networks from Inria are involved in different scientific challenges organised as part of TAILOR (WP 3, WP4, WP5, WP6, WP7). “I’m in charge of one of these cross-sectoral work packages, WP2, which has two components to it”, explains Marc Schœnauer. “Firstly, there’s the creation of a strategic roadmap, the aim of which is to prioritise research, training and technology transfer initiatives at a European level. The second component is concerned with organising challenges on the Codalab platform, coordinated by Isabelle Guyon from the TAU project team. These competitions, hackathons, and so on provide an important way of determining which algorithm is best-suited to solving a given problem.”
Four European networks for research excellence in AI
Alongside ELISE, HumanE-AI and AI4Media, TAILOR is one of four European networks to receive funding within the context of the Horizon 2020 programme Towards a vibrant European network of AI excellence centres. Its aim is to enable Europe “to boost its existing research capabilities [...] while promoting cooperation between leading research teams in order to find more effective ways of tackling the major scientific and technological challenges standing in the way of the deployment of AI-based solutions.” The overarching ambition is for Europe to guarantee its strategic autonomy “in AI, a critical field from a technological perspective”.
Marc Schoenauer, the architect behind the TAU project team
Marc Schoenauer has been an Inria director of research since 2001. A graduate of the École Normale Supérieure, he began his career at the CNRS, at the École Polytechnique’s Centre for Applied Mathematics. Since the late 1980s his research has focused on artificial intelligence, at the frontier between stochastic optimisation and machine learning. President of the AFIA (the French Association for Artificial Intelligence) from 2002 to 2004, Marc Schoenauer has sat on the editorial board for a number of scientific journals, and was involved in the Villani programme on AI.
In 2003 he created TAO (Thème Apprentissage et Optimisation - Theme Learning and Optimisation), a team which would go on to become TAU (TAckling the Underspecified) in July 2019, a joint project-team Inria, CNRS and Université Paris-Saclay within the Laboratoire Interdisciplinaire des Sciences du Numérique (LISN). TAU's research programme is focused on machine learning and is centred around four main areas of research: ethical and robust AI, combining reasoning and learning approaches, autonomous learning and organising challenges. The work carried out by the team is applied to three main fields: Computational Social Science, Energy Management and Data-based Digital Modelling.