Week 3 — Deep Learning

Held in winter 2026–27 — dates TBD. Prerequisite: Week 1 and scientific programming experience.

Week 3 turns to deep learning. The goal: by the end of the week you can architect, build, train, and tune deep learning models for your own research problems — and validate them so the results hold up, with cross-validation schemes that fit your data and pipelines free of data leakage. This is the core that Week 4 builds on when we turn to foundation models.

Objectives

By the end of the week you should be able to:

• Understand how deep learning models work — architectures, training dynamics, and where they fit in Earth systems science.
• Architect a model for a scientific problem: choosing structures, inputs, and outputs that match the science.
• Build and train deep learning models on your own data.
• Tune models systematically — hyperparameters, regularization, and diagnosing what's going wrong when training misbehaves.
• Design cross-validation schemes that match your data's structure — including the spatial and temporal dependence common in Earth systems data, where naive random splits overstate skill.
• Prevent data leakage across the whole pipeline — train/test contamination, leakage through preprocessing and normalization, temporal and spatial leakage, and target leakage — and recognize the too-good-to-be-true results it produces.

The week at a glance

The day-by-day schedule will be posted when the winter dates are announced.

The Week's core:

Deep learning fundamentals you build yourself — including the validation discipline of leakage-free splits and honest cross-validation — so that when you work with foundation models in Week 4, you understand what is happening under the hood and can trust what your models tell you.