Trainable Gaussian-based activation functions for sensor-based human activity recognition

Neural networks’ capability to model non-linear relationships strongly depends on their activation functions (AFs). This dependency makes the search for AFs with better performance a relevant open research subject. This research work proposes two Trainable Gaussian-based AFs for Multilayer Perceptron neural networks on sensor-based human activity recognition (HAR), namely, the Four-Parameter Activation Gaussian Radial Basis Function (T4GRBF) and the Weighted Gaussian Radial Basis Function (WGRBF). The T4GRBF considers the training of four parameters. Two of these parameters regulate the center and spread of the Gaussian shape separately. Another parameter regulates the center and shape simultaneously, and the last parameter regulates the output range of the AF. The WGRBF adjusts the Gaussian shape with only two trainable parameters, and it is then scaled by the AF input. The proposals are validated through experiments on the Opportunity and UniMiB SHAR benchmark datasets. The results found for the Opportunity dataset in this research work evidence that the sliding window segmentation method has a high impact on the AFs’ performance, and the results in the UniMiB dataset show that the trainable GRBF-based AFs improve the HAR models performance. Furthermore, in both datasets, the Trainable Gaussian-based AFs fit the training data better than the standard AFs, regardless of the specific time window setup used in the experiments.