Research Projects

Deep Neural Networks (DNN) have been widely used in various domains, including healthcare, activity recognition, and human behavior understanding. However, the challenge lies in the need for labelled data, as manual annotation is time-consuming, fatigue-inducing, and error-prone. With the rise of ubiquitous devices, most data is unlabelled, making it difficult to fully exploit the potential of this rich dataset due to significant annotation overhead. To address this issue, an annotation framework called HEXR is being developed to label time series signals with minimal expert intervention. The framework comprises self-supervised and semi-supervised approaches, a contrastive sampling method with a novel reconstruction loss, a parameterized activation function for optimization, and higher-order optimization using a smooth binary Quantile classification framework (SBQC). The project aims to evaluate HEXR in various domains, including mental wellbeing and daily activity. It will use multiple physiological signals (e.g., ECG, EEG, EDA) to annotate different mental states like stress and depression. Additionally, the approach will be tested in daily living activities such as surface identification for wheelchair users. The findings from these large-scale, diverse datasets may open the possibility of using HEXR framework for annotating large-scale datasets for other domains. By leveraging apriori information from domain experts and the intrinsic properties of dataset clusters, intelligent annotation approaches can be devised that significantly reduce human engagement.