Bonpagna KANN

Abstract:
The increasing use of smart devices, such as smartphones and smartwatches, has increased the availability of data collected from users in their daily activities. This data, gathered by integrated sensors, enables the training of machine learning models for human activity recognition (HAR) and the development of intelligent systems for health monitoring and diagnostics, as well as personalized services. Despite this progress, machine learning models trained on continuous data streams face a major problem: catastrophic forgetting. This problem arises when models fail to retain previously learned activities while trying to adapt to new activities introduced by users. To address this issue, continuous learning (CL) has emerged as a solution, allowing models to learn from evolving data streams while minimizing the forgetting of prior knowledge. Among the proposed approaches, replay-based methods are recognized as promising for maintaining the accuracy of past knowledge. However, their success hinges on balancing memory constraints and the quality of the samples selected for rereading. This thesis addresses this issue by proposing two generative rereading approaches and an experimental protocol within the framework of incremental class-based learning in HAR. First, we present TaskVAE, a memory-efficient approach that trains a lightweight, task-specific variational autoencoder (VAE) to generate synthetic data for past activities. TaskVAE has been shown to significantly improve CL classifier accuracy compared to experience replay methods, with an equivalent memory budget. Identifying the quality of the generated samples as a key factor, we then developed TaskVAE-GMM, an improved method that incorporates Gaussian mixture models (GMMs) to structure the latent space of the VAE during the sample generation process. By sampling from density-based clusters, TaskVAE-GMM produces higher-quality synthetic data, enabling more efficient retention of knowledge from previously learned tasks. Through extensive experiments using the proposed evaluation protocol, TaskVAE-GMM outperforms existing generative replay and experience replay methods across all domains, making it a robust and efficient solution for building adaptive and scalable HAR systems capable of incremental learning in real-world environments.