Abstract:

With data size too large to store on personal machines and processing time too long for a single GPU, Machine Learning (ML) workflow is turning to a new era of cloud-based approaches. In this report, we would introduce container technology and previous methods for managing GPU clusters for ML workflow, as well as the benefits and trade-offs of adopting container and its orchestration platform, Kubernetes. In addition, we would review an active field of Visualization for Machine Learning (Vis4ML), specifically interactive visual model analysis. We would describe several visual systems for general ML operation and demo a web application for easy model selection and inference, focusing on image classification and segmentation. Lastly, we would conclude with recommendations and future work.