Continual Symbol Aggregation: Aggregating Patches or Concepts for Continual Learning

Artificial Intelligence (AI) has achieved remarkable milestones in recent years in image recognition, computer vision, and medical diagnostics. However, most of these advancements rely on deep learning models that, despite their high accuracy, suffer from a significant lack of interpretability and continuous adaptability. This gap becomes critical in high-stakes medical applications, where models must evolve while providing interpretable results. One potential solution is to integrate symbol-based learning techniques within deep learning models. In this context, symbols represent interpretable information units that can express complex concepts or structures in the data. In images, symbols can be defined in various ways, depending on the representation and level of abstraction chosen as patches or concepts. Furthermore, this thesis explores this solution in continual learning, solving relevant issues such as privacy or reasoning shortcuts.