Chairs | Core elements of AI

• Traceable knowledge representation Ontology-based pruning and specialization 
• Certified AI algorithms and data security 
• Normalization and future legislation of AI 

• Small data, active learning, approximate methods
• Distributed and federated AI/edge AI 
• Online/real-time learning and decision 
• Reasoning under and against uncertainty

Learning and reasoning with constraints

Learning and inference are traditionally regarded as the two opposite, yet complementary and puzzling components of intelligence. In the last few years, Prof. Gori has been carrying out research on constrained based
models of the environmental agent interactions, with the main purpose of unifying learning, inference, and reasoning within the same mathematical framework. The unification is based on the abstract notion of constraint, which provides a representation of knowledge granules, gained from the interaction with the environment, as well as of supervised examples. The theory offers a natural bridge between the formalization of knowledge, expressed by logic formalisms, and the inductive acquisition of concepts from data.

Generative models for unsupervised and deep learning with complex data (2019-2023 / 2023-2027)

We focus on learning problems that are made difficult by real-world constraints, such as unsupervised deep learning, choosing a deep architecture for a given situation, learning from heterogeneous data or in ultra-high-dimensional scenarios. We seek to develop deep generative models, encoding sparsity priors, to address those issues. 

AI and natural language (2021-2025)

The goal of my research is to design debating technologies for advanced decision support systems, to support the exchange of information and opinions in different domains (as healthcare and politics), leveraging interdisciplinarity and advances in machine learning for Natural Language Processing.

Combining artificial and augmented intelligence technics on and through the web (2019-2023 / 2023-2027) 

Formalizing knowledge-based models and designing algorithms to manage interactions between different forms of artificial intelligence (e.g. rule-based, connectionist, and evolutionary) and natural intelligences (e.g. individual user, and crowd) on the web. 

Machine Learning for Computer Simulation (2021-2024)

Computer simulation has been widely used for planning high-impact decision-making (e.g., policies on climate change and Covid-19), but its reliability depends on how accurate simulations can imitate reality. This project develops machine learning methods to improve a simulator’s reliability and the resulting decision-making. 

Interpretability and security of statistical learning in healthcare (2019-2023 / 2023-2027)

Statistical learning in healthcare must ensure interpretability and compliance with secured data access. To tackle this problem, I will focus on 1) interpretable biomedical data modeling via probabilistic inference of dynamical systems, and 2) variational inference in federated learning for the modeling of multicentric brain imaging and genetics data.

Deep learning for dirty data: a statistical perspective (2021-2024)

The successes of machine learning remain limited to clean and curated data sets. By contrast, real-world data are generally much messier. We work on designing new machine learning models that can deal with “dirty” data sets that may contain missing values, anomalies, or may not be properly normalised. Collaborators include doctors and astronomers.

Neural network sparsity and applications (2024-2028)

We investigate interdisciplinary challenges spanning machine learning, computational neuroscience, and theoretical computer science. Our research addresses neural network sparsification, modeling brain organization, and understanding multi-agent systems. We develop algorithmic tools and mathematical frameworks to explore these areas, including the Assembly Calculus for cognitive processes and computational dynamics for distributed systems. We also apply our methodologies to theoretical biology, studying collective behaviors in biological systems.

PERUSALS: Pervasive Sustainable Learning Systems (2021-2025)

PERUSALS (Pervasive Sustainable Learning Systems) seeks to identify design principles of Internet-scale distributed learning systems, with a focus on the tradeoff between performance (in particular training and inference times), economic and environmental costs, and privacy.

Large Language Models for structured data (2024-2028)

We focus on addressing the limitations of Large Language Models (LLMs) in handling structured data, which hinders their adoption for data-centric tasks like analyzing large tabular datasets. We aim to develop novel frameworks and architectures for encoding structured data into LLMs, thereby enabling complex table understanding capabilities in pre-trained models.

Geometric statistics and geometric subspace learning (2019-2023 / 2023-2027)

We study the impact of topology (singularities) and geometry (non-linearity) of the data and model spaces on statistical learning, with applications to computational anatomy and the life sciences. The tenet is that geometry is critical when learning with limited resources and real-world constraints such as small data and limited computational resources.

Decision intelligence (2019-2023 / 2023-2027)

We are designing explainable decision-making processes satisfying real world constraints in a multi-objective environment including incomplete, fuzzy or stochastic data. 

BOGL: Bilevel optimization for graph learning (2024-2028)

The goal of the project BOGL is to bring together the methodology of bilevel optimization and graph machine learning. We seek to develop new algorithms to handle task on graphs such as graph classification, link prediction, community detection, taking into account the non-euclidean aspects of the data.

Finding structures in heterogeneous data (2022-2026)

We study heterogeneous data both in their kind and in their distribution. Our ambition is to discover structures in the data that will help the user in understanding it and taking decisions. We focus on designing generative models able to reveal several clustering viewpoints and we will adapt them to the deep-learning setting. We collaborate with doctors and retailers.

Artificial argumentation for humans (2019-2023 / 2023-2027)

The goal of my research is to design and create intelligent machines with the ability to communicate with, collaborate with, and augment people more effectively. To achieve this challenging goal, intelligent machines need to understand human language, emotions, intentions, behaviors, interact at multiple scales, and be able to explain their decisions. 

TopMoDaL: Multiparameter topological data analysis for Machine Learning Models and data sets (2024-2028)

The central tenet of the TopMoDaL project is that multiparameter topological data analysis (mTDA) has the potential to become an important asset for most standard machine learning (ML) models and pipelines. The aim of the project is to drastically improve the predictive and/or generative powers of ML models by providing both new descriptors and new regularization and monitoring tools from mTDA, that can be applied on many types of complex data.

Integrating AI into Network Solutions:  Exploring Privacy, Security, and Energy Efficiency (2024-2028)

One of the main objectives of my research project for the next few years is to explore how artificial intelligence techniques can be used to revisit classical and current network problems (related to Axis 1 – Fundamentals of AI). I have a special interest for privacy and security, in particular of federated learning models (related to the emerging topic AI for cybersecurity), and for the study of energy efficiency and frugality of information systems (related to the emerging topic AI for the environment).