Wearable sensors are widely used in activity recognition (AR) tasks with broad applicability in health and well-being, sports, geriatric care, etc. Deep learning (DL) has been at the forefront of progress in activity classification with wearable sensors. However, most state-of-the-art DL models used for AR are trained to discriminate different activity classes at high accuracy, not considering the confidence calibration of predictive output of those models. This results in probabilistic estimates that might not capture the true likelihood and is thus unreliable. In practice, it tends to produce overconfident estimates. In this paper, the problem is addressed by proposing deep time ensembles, a novel ensembling method capable of producing calibrated confidence estimates from neural network architectures. In particular, the method trains an ensemble of network models with temporal sequences extracted by varying the window size over the input time series and averaging the predictive output. The method is evaluated on four different benchmark HAR datasets and three different neural network architectures. Across all the datasets and architectures, our method shows an improvement in calibration by reducing the expected calibration error (ECE)by at least 40%, thereby providing superior likelihood estimates. In addition to providing reliable predictions our method also outperforms the state-of-the-art classification results in the WISDM, UCI HAR, and PAMAP2 datasets and performs as good as the state-of-the-art in the Skoda dataset.

Temporal patterns are encoded within the timeseries data, and neural networks, with their unique feature extraction ability, process those patterns to provide a better predictive response. Ensembles of neural networks have proven to be very effective Human Activity Recognition (HAR) tasks with time-series data, e.g., wearable sensors. The combination of predictions coming from the individual models in the ensemble helps boost the overall classification metric through efficient temporal pattern recognition. Currently, the most common strategy for combining the predictions coming from the individual models is simple averaging. However, since each ensemble model learns different temporal patterns of the timeseries classification problem, a simple averaging strategy is sub-optimal. This sub-optimality is addressed in this paper through a neural network-based adaptive learning framework. The method's core is training a neural gate that ingests the same input time-series data fed to the other temporal models. The goal of the training process is to adaptively learn scaler values against each temporal model by looking at the input data. These scaler values weigh each temporal model while combining the ensemble. The framework obtains superior predictive performance as compared to the standard ensembling techniques. The framework is evaluated on a benchmark HAR dataset called PAMAP2 [3] with two popular state-of-the-art ensemble architectures namely DTE [1] and LSTM-ensemble [2]. In both cases, the classification performance of the framework in HAR tasks surpasses the state-of-the-art models.

With the growing interest of the research community in making deep learning (DL) robust and reliable, detecting out-of-distribution (OOD) data has become critical. Detecting OOD inputs during test/prediction allows the model to account for discriminative features unknown to the model. This capability increases the model's reliability since this model provides a class prediction solely at incoming data similar to the training one. OOD detection is well established in computer vision problems. However, it remains relatively under-explored in other domains such as time series (i.e., Human Activity Recognition (HAR)). Since uncertainty has been a critical driver for OOD in vision-based models, the same component has proven effective in time-series applications.We plan to address the OOD detection problem in HAR with time-series data in this work. To test the capability of the proposed method, we define different types of OOD for HAR that arise from realistic scenarios. We apply an ensemble-based temporal learning framework that incorporates uncertainty and detects OOD for the defined HAR workloads. In particular, we extract OODs from popular benchmark HAR datasets and use the framework to separate those OODs from the indistribution (ID) data. Across all the datasets, the ensemble framework outperformed the traditional deep-learning method (our baseline) on the OOD detection task. I. INTRODUCTIONDeep learning (DL) methods for HAR are integral to many ubiquitous applications. E.g., providing live coaching feedback to an athlete based on mobile or on-body sensor data requires an efficient HAR algorithm. The algorithm predicts if the person is running, walking, jogging, etc., and coaching feedback is generated based on that prediction. In such applications, it is common to encounter unseen out-ofdistribution (OOD) activities with respect to known or indistribution (ID) activities. E.g., taking a rest while running or performing some spontaneous activity such as taking a phone call. The model does not know the above activities (hence OOD). Therefore, it must differentiate the OOD data from ID data in those scenarios. Failing to do so leads to misclassification, affecting model reliability. However, most state-of-the-art DL models used for HAR fail to do so. The primary reason is that these models are trained to discriminate between classes with high accuracy without *This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 813162. The content of this paper reflects the views only of their author (s). The European Commission/ Research Executive Agency are not responsible for any use that may be made of the information it contains.