Explicações visuais aplicadas a redes neurais convolucionais unidimensionais com ênfase no reconhecimento humano de atividades

Abstract

Human Activity Recognition (HAR) is an area of growing importance, especially with the popularization of wearable devices. One-dimensional convolutional neural networks (1D CNNs) have emerged as a widely used approach for HAR. These architectures are data-driven and have the ability to learn complex patterns in signals that can be used to classify human activities. However, although 1D CNNs can achieve impressive numerical results, understanding and explaining the decision-making of these models remains a challenge. This work proposes two innovative approaches to generate visual explanations in 1D CNNs applied to HAR, the first one utilizing Gradient-weighted Class Activation Mapping (grad-CAM) and the second one utilizing t-Distributed Stochastic Neighbor Embeddings (t-SNE). Our goal is, through visualizations, to understand and explain the complex patterns learned by 1D CNN models during the training process, proposing visualizations that enable the understanding of the models' decision-making process. The proposed explanations allow the identification of biases in the models and datasets, analysis of how the validation approach impacts model learning, and also the selection of a better model by observing not only numerical but also qualitative results. In general, based on the experiments proposed in this work, the combination of explainable artificial intelligence techniques with deep learning has the potential to provide a holistic view of the capabilities of trained models, making it possible to create explanations and formulate hypotheses about the models' decision-making process.