Tutorial on ”Machine Learning Lifecycle in High Performance Computers and Cloud: A Focus on Geospatial Foundation Models”

General Information

Foundation models (FMs) are pre-trained neural networks used for various computer vision tasks such as object classification, detection, and segmentation. These models are trained on large, diverse datasets, enabling them to capture a wide range of visual features applicable across different domains, making them highly versatile. The key advantage of FMs is their ability to perform tasks without requiring extensive datasets for training from scratch. By leveraging deep learning’s representational capabilities, FMs can generalize across tasks, improving efficiency and accessibility for various applications. Their development is a collaborative effort involving interdisciplinary teams, promoting open science principles like transparency and knowledge sharing. In this context, hybrid approaches are necessary for the main training of FMs using High-Performance Computing (HPC), while fine-tuning and inference with new data can be conducted in a cloud computing environment.

The tutorial will begin by covering recent advancements in High-Performance Computing (HPC) and Cloud Computing, as well as the basics of Foundation Models (FMs). Participants will learn how HPC can support the development of large-scale FMs, and how these models can be deployed in the cloud for public use. In the hands-on sections, participants will access AWS Cloud resources and learn how to fine-tune FMs for selected Earth Observation downstream tasks.

Instructors

Gabriele Cavallaro – Forschungszentrum Jülich

 Rocco Sedona –  Forschungszentrum Jülich

Manil Maskey – NASA

Iksha Gurung – University of Alabama in Huntsville

Sujit Roy – NASA

Muthukumaran Ramasubramanian – NASA

Rahul Ramachandran – NASA

Rajat Shinde – NASA

Teaching Material

• NASA-IMPACT/HDCRS-school-2024