Machine Learning

Our work in machine learning spans theory, algorithms and applications. We have been studying different learning settings, including online learning, learning with partial feedback, learning in games, unsupervised learning, meta-learning, and statistical learning theory. Researchers in the unit have made significant contributions to the foundations of machine learning, leveraging tools from statistics and probability, optimization, and theoretical computer science. Underpinning the theory of machine learning is important in order to understand the success and failure of learning algorithms, and plays a key role in order to make machine learning a practical tool and reliable.


We are interested in applying machine learning to develop robots that can efficiently adapt to the surrounding environment and solve complex tasks without or with minimal manual programming. We study algorithms that allow robots to learn on-line with limited or without human supervision as well as active learning strategies that can guide the robot in the exploration of the environment to speed up learning. We exploit the rich sensory system available on humanoid robots to study learning algorithms that leverage on multi-modal information (e.g. vision and touch) and dynamic information from neuromorphic sensors. Applications include: object manipulation, scene and object perception, locomotion, social as well as physical human-robot interaction

Natural Intelligence

Our research investigates how biological brains make decisions leveraging many sources of small information, both in the sensory information processing and in the social reasoning domain. Understanding these natural computations is key both to translate principles of natural intelligence into artificial intelligence and to make an impact in healthcare of brain diseases. We build algorithms that unite observation and perturbation of large scale neural activity, to uncover the principles of how circuits of neurons make decisions during complex behavior. Through interdisciplinary collaborative efforts, these algorithms are used to reveal the principles of how populations of neurons in different brain areas work together to accumulate and amplify weak evidence for correct decision making on naturalistic sensory signals, and uncover the neural computations underlying the processing of complex social signals such as intention communications from body kinematics. We also work to translate the information coding principles found in natural neural sensory networks to robotic sensors.

Machine Learning for Health

The availability of unprecedented amount of digital data, combined with advances of machine learning, is at the basis of predictive, personalised, preventive and precision medicine. By leveraging on prior medical knowledge and doctors expertise, we study algorithms that can lead to early diagnosis and more effective therapy for a wide range of diseases. The many challenges we face include the collection and the analysis of data fully respecting privacy and protection concerns, the development of explainable methods, and the design of algorithms able to integrate highly heterogeneous and time dependent data. We devise algorithms for the extraction of information from health databases, both at the regional and national level. We consider different data scenarios, including small datasets (rare diseases), large datasets (mass screening), as well as very high dimensional datasets (whole genome population). We are also engaged in the development of data driven structural models able to advance the understanding of disease pathogenesis and of the underlying biological mechanisms and pathways.
© Ellis Unit Genova