Foundation Models at the Edge
Foundation models, starting with BERT for natural language processing, have vastly outperformed prior approaches in the last few years. These models exploit vast volumes of unlabeled data using self-supervision and produce base models that can be adapted to a wide range of downstream tasks. More recently, their adoption in non-NLP domains, particularly, remote sensing data has received significant attention. This talk will revisit AI/ML model lifecycle with a foundation model lens and present key opportunities and challenges in edge deployments (e.g., Kubernetes cluster in space). Such edge deployments introduce at least two types of challenges: (i) resource constraint at the infrastructure layer, and (ii) lack of human supervision at the data and AI layer.
These problems are seemingly exacerbated due to foundation models, which are generally larger than their predecessors (statistical or deep learning models). However, foundation models also create new opportunities which can make them more attractive for operationalization. The ability to learn common representation across multiple downstream tasks allows downstream tasks to share common subnets (reducing memory footprint and model loading time). Representations learnt by foundation models become a key input to data retention (e.g., unusual embedding vector => higher retention) and data selection for human supervision such as labeling and model re-training (e.g., select diverse embedding vectors for human supervision). This talk will outline some of these challenges and opportunities in operationalization of foundation models at the Edge.