Our research is organized as follows:
|Resource Efficient Machine Learning
|Foundations of Learning Theory
|Optimization for Machine Learning
|Structured Machine Learning
|Explainable and Reproducible Machine Learning
|ML for Physics
|Machine Learning for Health
Resource Efficient Machine Learning: The availability of large scale datasets requires the development of ever more efficient machine learning solutions.
A key feature towards scalability is to tailor computational requirements to the generalization properties in the data, rather than their raw amount.
This project aims at blending statistical and algorithmic principles to design new, sound and scalable learning machines.
Optimization for Machine Learning: Numerical aspects are increasingly central in machine learning.
The size and diversity of data and corresponding machine learning problems require flexible and scalable optimization solutions.
We aim at developing sound optimization algorithms tailored to tackle a variety of modern machine learning problems. Our approach puts emphasis on robustness and adaptivity to the underlying problem geometry, which is exploited to derive provably flexible and efficient solutions.
Structured Machine Learning:
Machine learning often deals with Euclidean data, however many modern applications require dealing with data naturally described by other structures, such as: graphs, strings and sequences, trees, curves, probability distributions, time series and dynamical systems. We tackle the problem of dealing with structured data from a theoretical, algorithmic and practical point of view. We consider a number of applied scenarios, including inferring structure in unstructured textual content from the web, identifying subgraphs based on prior knowledge for recommender systems, forecasting the evolution of time-dependent phenomena.
Explainable and Reproducible Machine Learning:
As machine learning is being applied in an ever increasing number of real-life scenarios, it is mandatory to design methods that are stable and provide reproducible results. Further, in many contexts, models should not only be predictive but also provide explainable and interpretable solutions. This is of particular importance in the common setting in which data are extremely sparse or heterogeneous. We work towards these goals using sparse models as well as integrating data modalities to capture complex interactions while ensuring explainability.
Machine Learning for Health:
Data availability and algorithmic advancements provide a unique opportunity to develop a new generation of health technologies. Machine learning provides tools to tackle a wide range of questions including for example early-detection and automatic non-invasive diagnosis of disease, accurate prediction of disease progression, identification of factors underlying pathogenesis. The large-p small-n scenario is paradigmatic in biological and medical applications. Appropriate machine learning solutions must be defined taking into account this peculiarity and the heterogeneity of the data at hand, which may include molecular-omics data, patient-reported outcomes, clinical records and imaging, to name a few. Our focus is largely on neurological, oncological and immunological diseases.
Foundations of Learning Theory: We work on the theoretical foundation of machine learning blending within a multidisciplinary approach combining probabilistic and numerical tools with an emphasis on analytic approaches.
Our view on machine learning highlights the connection with other fields such as inverse problems and regularization theory, signal processing and harmonic analysis.
Research in this direction is done in collaboration with CHARML@MalGA.
Vision: In recent years, machine learning approaches have revolutionized computer vision.
While current solutions provide remarkable performances, they are often data hungry and computational expensive.
We work with computer vision researchers to develop machine learning solutions for vision that are efficient both with respect to human supervision and computation. Research in this direction is done in collaboration with MLV@MalGA.
Robotics: Robotics is a natural testbed for machine learning solutions.
The variety of sensory modalities robots are endowed with requires for the robot learning to adapt and interact with the environment and humans.
We develop and apply cutting edge machine learning techniques to solve perception, cognition and control problems in humanoid robotics.
Research in this direction is done in collaboration with HSP@IIT and DIC@IIT.
ML for Physics: We work on developing machine learning tools for the efficient and accurate analysis and modelling of high energy physics (HEP) and meteorological data, as well as the study of navigation problems.
Research in this direction is done in collaboration with ML4DS@MaLGa.