I spend a portion of every week teaching neural networks to folks who have never trained one before. The classes feel calmer when we start with tiny, debuggable models.
Why small models matter
Compact models are interactive. Students can inspect every weight, rerun training in seconds, and connect loss curves to real intuition.
- Inspectable activations lower anxiety.
- Fast iteration keeps curiosity alive.
- Edge cases become conversation starters instead of blockers.
A recipe for the first lesson
- Start with a simple dataset—think 200 rows, no hidden columns.
- Train a logistic regression model in a notebook while narrating every decision.
- Invite participants to change one feature at a time and observe outcomes.
Teaching is slower than shipping. The pace is the point.
Scaling up intentionally
Once confidence is high, I introduce modern architectures with the same care: show the simplest working example, highlight new vocabulary, and return to projects that feel meaningful to the audience.