Improving Human Activity Recognition Models by Learnable Sparse Wavelet Layer

Modern machine learning algorithms for human activity recognition based on artificial neural networks often require a large amount of labelled training data to generalize between human subjects and training contexts. Large degrees of freedom make them susceptible to overfitting and often computationally intensive to implement on portable hardware. In this work, we introduce wavelet-based learnable filters as a feature extraction layer that greatly improves the generalization capability of the detector model. Our evaluations on six benchmark datasets show significant improvements in macro $F_1$ score when our wavelet-based learnable filter layer is prepended to three state-of-the-art human activity recognition models. As a side effect, in many cases we could drastically reduce the required model size to achieve competitive performance on the benchmark dataset, which is an important requirement for use in wearable computing.