Towards Self-Supervised Learning @ the Edge

Dario Dematties, Northwestern-Argonne Institute of Science and Engineering (NAISE)
LANS Seminar Graphic featuring the title and date for the event.

In the realm of deploying machine learning (ML) algorithms on edge devices, a plethora of challenges surfaces, ranging from compute resource limitations to the nuanced intricacies associated with acquiring dynamic data distributions. The majority of extant ML algorithms implemented on edge cyberinfrastructures undergo training with datasets that inadequately capture the true essence of the data streams collected at the edge. This presentation delves into an investigation of the viability of novel self-supervised learning algorithms to effectively characterize insufficiently curated, imbalanced, and unlabeled datasets in the context of edge computing. The primary focus is directed towards their potential applicability in addressing the unique challenges posed by edge environments. Two specific applications will be presented in this talk: Atmospheric Conditions Characterizations and Ecosystem Activity Monitoring. Both instances exemplify the autonomous characterization capabilities facilitated by the proposed self-supervised learning algorithms, obviating the need for human intervention in the process. Through these case studies, insights into the transformative potential of self-supervised learning in mitigating the challenges associated with deploying ML algorithms on edge devices will be addressed.

Bio: Dario is a Computer Vision/Deep Learning Postdoctoral Researcher at Northwestern‑Argonne Institute of Science and Engineering (NAISE) with expertise in developing self-supervised learning models and deploying them efficiently on edge devices and high-performance computing systems. He has leveraged techniques like contrastive learning, vision transformers, and joint embeddings for applications such as atmospheric conditions characterization and ecosystem monitoring. His research interests include the combination of self-supervised learning with reinforcement learning to enable truly autonomous data acquisition through intelligent, embodied agents. He has experience in federated learning and model optimization.  Dario received his Ph.D. from University of Buenos Aires, Argentina in 2020 and worked as a postdoctoral researcher at the National Scientific and Technical Research Council (CONICET) in Argentina before joining NAISE. 

See all upcoming talks at