<aside> 🧭

The goal of the homework is to observe, analyze, and explain trustworthy AI problems, rather than implement complex algorithms from scratch. The assignments are intentionally open-ended: you may choose appropriate models, datasets, APIs, open-source tools, or existing implementations. Properly acknowledge any external resources you use.

</aside>

Students enrolled in DATA 405 complete the core requirements. Students enrolled in DATA 603 additionally complete the Graduate Extension for each homework.

Unless otherwise specified, submit a concise 1–2 page report. Figures, tables, references, and code are not counted toward the page limit. Include code or links only when code is part of your investigation.

Homework 1: Generalization — What Happens When the Data Changes?

Machine learning models may perform well on familiar data but fail when the test distribution changes. Your goal is to demonstrate and analyze one simple example of distribution shift.

What to Do

Choose any model, dataset, and setting that you are comfortable with. You may train a small model, use a pretrained model, reuse an existing open-source implementation, or adapt an experiment from another source.

Create or identify at least two test conditions: one closer to the original distribution and one containing a meaningful distribution shift.

Possible examples include:

• clean vs. corrupted images,
• one visual domain vs. another,
• different populations or environments,
• different writing styles or prompt formats,
• or another meaningful distribution change.

What to Submit

1. Briefly describe the model, data, and distribution shift.
2. Compare model performance before and after the shift.
3. Include at least one table or figure showing the difference.
4. Explain why you think the performance changed.