Deep Learning models have achieved remarkable performance in tasks such as image classification or generation, often surpassing human accuracy. However, they can struggle to learn new tasks and update their knowledge without access to previous data, leading to a significant loss of accuracy known as Catastrophic Forgetting (CF). This phenomenon was first observed by McCloskey and Cohen in 1989 and remains an active research topic. Incremental learning without forgetting is widely recognized as a crucial aspect in building better AI systems, as it allows models to adapt to new tasks without losing the ability to perform previously learned ones. This article surveys recent studies that tackle CF in modern Deep Learning models that use gradient descent as their learning algorithm. Although several solutions have been proposed, a definitive solution or consensus on assessing CF is yet to be established. The article provides a comprehensive review of recent solutions, proposes a taxonomy to organize them, and identifies research gaps in this area.

In this article, we introduce explainable methods to understand how Human Activity Recognition (HAR) mobile systems perform based on the chosen validation strategies. Our results introduce a new way to discover potential bias problems that overestimate the prediction accuracy of an algorithm because of the inappropriate choice of validation methodology. We show how the SHAP (Shapley additive explanations) framework, used in literature to explain the predictions of any machine learning model, presents itself as a tool that can provide graphical insights into how human activity recognition models achieve their results. Now it is possible to analyze which features are important to a HAR system in each validation methodology in a simplified way. We not only demonstrate that the validation procedure k-folds cross-validation (k-CV), used in most works to evaluate the expected error in a HAR system, can overestimate by about 13% the prediction accuracy in three public datasets but also choose a different feature set when compared with the universal model. Combining explainable methods with machine learning algorithms has the potential to help new researchers look inside the decisions of the machine learning algorithms, avoiding most times the overestimation of prediction accuracy, understanding relations between features, and finding bias before deploying the system in real-world scenarios.

Smartphones have emerged as a revolutionary technology for monitoring everyday life, and they have played an important role in Human Activity Recognition (HAR) due to its ubiquity. The sensors embedded in these devices allows recognizing human behaviors using machine learning techniques. However, not all solutions are feasible for implementation in smartphones, mainly because of its high computational cost. In this context, the proposed method, called HAR-SR, introduces information theory quantifiers as new features extracted from sensors data to create simple activity classification models, increasing in this way the efficiency in terms of computational cost. Three public databases (SHOAIB, UCI, WISDM) are used in the evaluation process. The results have shown that HAR-SR can classify activities with 93% accuracy when using a leave-one-subject-out cross-validation procedure (LOSO).