This paper explores the performance of deep, convolutional, and recurrent models for human activity recognition (HAR) using wearable sensors. The study evaluates three representative datasets containing movement data captured with wearable sensors. The authors describe how to train recurrent approaches, introduce a novel regularization method, and demonstrate that these models outperform the state-of-the-art on a large benchmark dataset. Across thousands of recognition experiments with randomly sampled model configurations, the study investigates the suitability of each model for different tasks in HAR, explores the impact of hyperparameters using the fANOVA framework, and provides guidelines for practitioners applying deep learning in their problem setting. The paper discusses the potential of deep learning to significantly impact HAR in ubiquitous computing. It highlights the challenges of traditional methods, such as manual feature extraction and classification, and the benefits of deep learning in handling complex behaviors. The study compares deep feed-forward networks (DNNs), convolutional networks (CNNs), and recurrent networks (including LSTM and bi-directional LSTM) in terms of performance and suitability for different HAR tasks. The results show that recurrent networks, particularly bi-directional LSTMs, outperform the state-of-the-art on the Opportunity dataset, a large benchmark dataset. The study also finds that CNNs are more suitable for prolonged and repetitive activities like walking or running. The paper provides insights into the impact of hyperparameters on performance and offers guidelines for practitioners to select and tune models for their specific HAR applications. The findings suggest that recurrent networks are well-suited for real-time applications due to their ability to model movement at the sample level. The study emphasizes the importance of parameter exploration and provides a systematic comparison of deep learning approaches for HAR.This paper explores the performance of deep, convolutional, and recurrent models for human activity recognition (HAR) using wearable sensors. The study evaluates three representative datasets containing movement data captured with wearable sensors. The authors describe how to train recurrent approaches, introduce a novel regularization method, and demonstrate that these models outperform the state-of-the-art on a large benchmark dataset. Across thousands of recognition experiments with randomly sampled model configurations, the study investigates the suitability of each model for different tasks in HAR, explores the impact of hyperparameters using the fANOVA framework, and provides guidelines for practitioners applying deep learning in their problem setting. The paper discusses the potential of deep learning to significantly impact HAR in ubiquitous computing. It highlights the challenges of traditional methods, such as manual feature extraction and classification, and the benefits of deep learning in handling complex behaviors. The study compares deep feed-forward networks (DNNs), convolutional networks (CNNs), and recurrent networks (including LSTM and bi-directional LSTM) in terms of performance and suitability for different HAR tasks. The results show that recurrent networks, particularly bi-directional LSTMs, outperform the state-of-the-art on the Opportunity dataset, a large benchmark dataset. The study also finds that CNNs are more suitable for prolonged and repetitive activities like walking or running. The paper provides insights into the impact of hyperparameters on performance and offers guidelines for practitioners to select and tune models for their specific HAR applications. The findings suggest that recurrent networks are well-suited for real-time applications due to their ability to model movement at the sample level. The study emphasizes the importance of parameter exploration and provides a systematic comparison of deep learning approaches for HAR.