Innovative Models for Machine Learning – IMML

Research Activity

The activity that this Research Group intends to carry out falls within the field of Machine Learning, with particular reference to the design and implementation of new models and to their introduction into the literature. Particular attention will be paid to the areas of Evolutionary Algorithms, Swarm Intelligence, and Artificial Neural Networks.

In this general area, there are several fundamental aspects of the research activity, which give rise to a set of specific activities indicated below.

The first activity consists of proposing innovative models of both Evolutionary Algorithms and Swarm Intelligence. Given the experience of the members of this Group, we are aware of both the strengths and weaknesses of the existing models. The aim of this activity is to overcome these limitations without, at the same time, losing the benefits through the introduction of improvement models.

A second activity consists in the design and use of Evolutionary Algorithms and Swarm Intelligence capable of allowing a representation of knowledge that is explicit and easily understandable by the user (Interpretable Machine Learning – IML). This activity goes in the direction, for example, of the most recent European Union directives in the medical field aimed at the “right to an explanation” for a patient: a user must be able to ask for the reasons for a decision made on himself through an algorithmic process. This activity will allow tackling problems in various areas such as IML-based classification on data sets, IML-based prediction and regression on signals, multi-variable optimization, even multi-objective and in the presence of constraints. Particularly fruitful areas are those of medicine, biology, healthcare and robotics.

A third activity concerns Federated Learning, a Machine Learning technique that allows learning from multiple sets of data stored on multiple decentralized devices or servers without the data itself having to travel. The design and implementation of innovative algorithms for Federated Learning is envisaged based on evolutionary or swarm intelligence algorithms.

A fourth activity consists in studying the use of Evolutionary Algorithms and Swarm Intelligence in the context of Neuroevolution, with particular reference to the optimization of the structure and the parameters of both ‘shallow’ and ‘deep’ neural networks. This is a largely felt problem within the community of network users, as demonstrated by studies on the subject undertaken, for example, by Google, Sentient Technologies, MIT Media Lab, Johns Hopkins, and Carnegie Mellon.

A fifth activity concerns the explainability of the solutions proposed by artificial neural networks. These latter are, in fact, ‘black boxes’, meaning that the solutions found do not provide clear explanations understandable to the user, and these must be found ‘a posteriori’. It is important to evaluate the characteristics of the already existing explainability methodologies so as to be able to identify the most interesting ones for the type of specific problem to be addressed from time to time.

A sixth activity is related to the design and implementation of parallel and distributed versions of Evolutionary Algorithms and Swarm Intelligence to reduce execution times and, at the same time, perform a better search in the solution space. In fact, one of the limitations of the current versions consists in their execution times, which often make their use in quasi-real time difficult.

Goals

The fundamental goal consists in the design, implementation and introduction into the literature of new Machine Learning models. Particular attention will be paid to the areas of Evolutionary Algorithms, Swarm Intelligence, and Artificial Neural Networks.

This general goal is expressed in a set of further specific goals: i) simple interpretability of the solutions proposed by evolutionary and Swarm Intelligence models; ii) possibility of learning in a federated manner; iii) support for connectionist deep learning by identifying suitable deep structures and the relative values ​​of their parameters that allow good performance; iv) effective explainability of the solutions proposed by the neural models; v) reduction in execution times so as to be able to achieve solutions in quasi-real time. Each of these goals gives rise to an activity briefly described above.

Application Fields

The methodologies in question are ‘general purpose’, so they may have a wide range of application fields. Among these latter, medicine, biology, healthcare, and robotics are of particular, although not exclusive, interest to our Research Group.