Grant PID2021-123090NB-I00 funded by: 
In recent years, big-tech companies and universities have invested a lot of resources to turn prototype applications based on the multidimensional and multiway analysis of user data into cutting-edge technology that impacts our living experience. These technological advances are based on novel learning techniques that seek to improve our health, productivity, and efficient use of energy.
Traditionally, many of the parametric methods in the field of statistical signal processing have been based on building generative models of observations. The recent success in deep learning has put forward a completely different paradigm, with over-parameterized representations trained on large amounts of supervised data, which guide search of the learning algorithms. Despite their advantages, the flexibility of these data-driven representations usually comes with the price of the lack of identifiability in the parameter space, which usually favors a learning based on correlations rather than the desirable causal relations. Moreover, the parameters usually lie in smooth manifolds whose Riemannian structure and underlying symmetries are usually not exploited by the learning algorithms.
In this project, we would like to combine model-based and data-driven approaches through the unfolding of the algorithms, with the aim to retain the best of these learning approaches, as well as to propose advances in representation methods based on unsupervised learning. We will apply these techniques to challenging problems, such as improving EEG signal processing techniques and brain-computer interfaces, exploring emotion recognition in interactive virtual environments, and researching distributed learning approaches for nonlinear behavioral models. In this sense, there is a consensus on the need for improved brain interfaces that enable more natural communication between users and devices without the need for keyboards. On the other hand, in the new generations of communications systems, the modeling and compensation of the nonlinearities introduced by the power amplifiers require distributed learning approaches with an optimized balance of computational complexity on the edge of the network.