Mejoras a la capacidad de generalización de la inteligencia artificial

Résumé

Information fusion is a crucial aspect of modern data analysis and decision making. It involves the integration of multiple sources of information in order to form a more complete and accurate understanding of a given subject. This process is particularly important in fields such as computer science, engineering, and natural sciences, where large amounts of data are generated from a variety of sources and must be synthesised to make informed decisions. Information fusion is also essential in the design and implementation of intelligent systems, as it allows the integration of various sensors and data sources to make more accurate predictions and recommendations. From a mathematical point of view, one way to study this problem is through the idea of fusion functions, which take as input a vector of numbers and return a single one, representative of them. A relevant kind of fusion function is the family of aggregation functions. These functions hold two boundary conditions and monotonicity with respect to the inputs, which induce some desirable properties to the function output. However, information fusion in applied systems comprises more than this theoretical notion. As the heterogeneity, the structure, and the volume of the data become more relevant, other approaches to tackle this problem have arisen. For example, in a network structure, the different inputs are associated among each other according to a pre-established set of relationships; in time series, data present temporal dependencies. When dealing with non-structured data, like text, audio, and image, deep learning approaches have been very successful in transforming this kind of data into vectorial representations of real numbers using series of affine transformations. Despite previous efforts in the field, the problem of effectively combining diverse and heterogeneous sources of information, remains an open and active area of research. This is due to the challenges inherent in integrating multiple sources that may be in different formats and may have conflicting or incomplete information. For example, how the information measured relates to other sources of data and how reliable those measures are is highly dependent on the measurement procedure. Indeed, systems that fuse the information from those different sources shall present additional complexities as well when taking into account the particularities of each feature considered. In this dissertation, we propose a collection of functions and algorithms to take into account possible interactions, heterogeneities, and uncertainties when working with different sources of information. We do so by means of aggregation theory and social network analysis, and we focus especially on those cases where deep learning approaches are not so successful. We apply these results to a wide range of problems, including the classification of brain computer interface signals, the classification of standard tabular data, the detection of anomalies, and the detection of communities in social networks.