arena-in-12-weeks
glossary about

format & ground rules

How the course works

ARENA in 12 Weeks repackages the official ARENA 3.0 exercises for a part-time study group. The target reader can write Python but has not studied ML.

Pair groups

Participants work in standing groups of 2 to 3, with no TAs. During the week, read the explainer. Once a week, meet for 3 to 4 hours and work through the official Colab notebook together (one shared screen, rotate who drives).

Ground rules

  1. The explainer is not optional; it supplies the context ARENA assumes.
  2. Finish core sections first. Stretch sections are bonus.
  3. Run everything. Code comprehension comes from executing and changing snippets.
  4. Use the solutions notebook when stuck, then re-derive the result in your own words.

Colab setup

Open the exercise notebook from the week page, then choose Runtime → Change runtime type → T4 GPU. Free Colab is enough for the planned core path.

FAQ

Do I need local setup?

No. The course assumes Google Colab for exercises. The site is static and has no accounts.

What if ARENA notebooks change?

Links point to ARENA's main branch for freshness. Each cohort should record the ARENA commit hash it starts from.

Why optimizers (week 3) before backprop (week 4)?

You train a network in week 2 treating backward() as a black box, then study optimizers (which act on gradients) in week 3, and only open the backprop box in week 4. The order is deliberate: use the training loop first, name its moving parts, then demystify where gradients come from. Treat weeks 3 and 4 as a paired unit rather than swapping them. Reordering would also break the week 2 → week 3 MNIST continuity.