Speaker: Stefano Cagnoni, UNIPR

Evolutionary algorithms (EAs) are a family of optimization techniques inspired by natural evolution. As stochastic optimization techniques that explore the solution search space in an extensive and parallel way without making use of the mathematical properties of the functions they optimize, EAs have some advantages over the gradient descent techniques used, for example, in deep learning. Firstly, they manage to avoid being “captured” by local minima, as often happens with gradient descent. Secondly, they use extremely flexible solution representations, which enable them to tackle problems of very different nature. In the course, problems encountered in machine learning applications will be considered, therefore proposing EAs as an alternative to machine learning techniques, compared to many of which they require a much smaller number of data for training and solution representations that are easier to interpret than the “black box” ones provided, for instance, by neural networks. These properties are fundamental in medicine, where studies are often characterized by a limited quantity of available data, besides requesting justifications of the answers provided by the automatic systems applied to them. The course aims to introduce the main variants of evolutionary algorithms (genetic algorithms, genetic programming, swarm intelligence) and demonstrate some application examples, involving the processing of signals and images of medical interest.