Superiority of Deep Learning in Ambient Intelligence, a Myth or an Ubiquitous Truth?

This research examines the effectiveness of Deep Learning (DL) and classical Machine Learning (ML) in human activity recognition within smart environments, showing that the performance difference between the two is often minimal. By critically analyzing documented DL performances and comparing them with our own ML models, we provide an empirical perspective on DL’s real usefulness in ambient intelligence. We reviewed papers in the field and selected common datasets for smart homes to compare classical ML methodologies with DL counterparts, focusing on hyperparameter tuning to optimize model performance. The hypothesis was that DL benefits are often overstated, and well-tuned classical ML methods can achieve competitive results more efficiently and simply. Our findings reveal the strengths and weaknesses of each approach, offering a critical perspective on their applicability in smart environments. For instance, a well-tuned SVM outperformed all approaches in the literature on the UCI-HAR dataset. Our findings can be perceived as quite surprising, suggesting to ambient intelligence practitioners to be critical in their technology choices.