Module 3 Instructor Notes

Teaching Goal

Students should connect optimization, loss, and regularization to concrete design choices, not only vocabulary.

Misconceptions

• Treating every deep model as a transformer

• Ignoring data scale and label quality

• Assuming GPU access fixes poor problem framing

Facilitation

• Start from a failure case, then introduce the method that addresses it.

• Keep one running example (e.g., image or text) across modules when possible.

Grading Cue

Reward clear reasoning about tradeoffs; do not require state-of-the-art benchmark scores in introductory work.