Optimization of Hyperparameters in Convolutional Neural Network for Human Activity Recognition

Human Activity Recognition (HAR) is a popular research area due to its diverse applications including smart health. Classification of various human activities mainly relies of manually extracted features based on domain knowledge and different machine learning classification methods. Sometimes it’s very tedious to extract the features manually from high voluminous sensor data. Thus, deep learning methods are gaining popularity due to its automatic feature extraction characteristics. In this research work, Convolutional Neural Network (CNN) is used to identify different human activities. However, tune the values of different hyperparameters associated with CNN to achieve higher classification accuracy is a challenging task. Expert experience with the combination of trial and error optimization is a common solution to tune the hyperparameters. Moreover, it is time consuming. In this work, we use Random Search (RS) to optimize the hyperparameters of CNN. In our experiment, standard public “Wireless Sensor Data Mining” (WISDM) dataset for HAR, is used to show the outperformance of our proposed method.