Grant: Discovery Grants Program - Individual

Human-generated signals, commonly captured via wearable sensors, smartphones, and smart home or vehicle systems, contain a broad range of behavioral, physiological, and cognitive insights. As a result, they can be widely used to enhance individual and societal well-being through applications such as physical and mental health, fitness and wellness management, assisted living and accessibility, and human-computer interaction. However, the inherent complexity and varied nature of different time-series data, along with the ever-changing circumstances of users, pose significant challenges in developing effective artificial intelligence (AI) systems and representation learning models. The primary objective of this research program is to devise novel deep learning architectures and techniques that can efficiently capture the underlying dynamics and interdependencies within and across different modalities of human- human-generated (a.k.a. human-centered) time-series data. More specifically, we will first design and develop advanced foundation models capable of understanding and predicting the ever-changing patterns inherent in multimodal time-series data. These models aim not only to classify current events but also forecast future events accurately, and automatically adapt to new data domains. These foundation models can be used to perform a variety of downstream tasks by capturing the rich temporal dynamics of multimodal human-centered time-series. Next, advanced capabilities will be designed and developed that enable foundation models to adapt to different users and scenarios. This necessitates the utilization of frameworks such as federated learning, continual learning, and contextual information learning. Moreover, considering the sensitivity of personalized AI, we will ensure that the designed representation learning frameworks are robust, fair, and interpretable through responsible AI principles. The proposed research program promises to introduce ground-breaking approaches in the area of deep representation learning for human-centered time-series, addressing key gaps in the current state of the art and advancing the field forward. Unlike traditional models that treat different modalities of human-centered time-series data in isolation, our approach seeks to uncover the underlying dynamics and interdependencies within and across these modalities, setting our research apart. We aim to harness the potential of large foundation models, continual learning of information, contextual grounding, and responsible AI principles in a unified manner, something that hasn't been approached holistically in previous work. The program will train 3 PhD, 4 MASc, and 3 Undergraduate students, equipping them with essential skills in the domains of deep learning, representation learning, time-series analysis, and responsible AI, and further contribute to the Canadian AI ecosystem.