Assignments

Three programming assignments — no separate project. Every assignment grows out of a notebook you’ve already seen in lecture — the work is never “build something alien”, it’s “take the thing we built together and push it until you own it”. Each covers only the material taught by its deadline, and each is followed by an individual post-deadline quiz/viva where you explain your own code, plots, and choices out loud. That conversation is where understanding is actually verified.

The three assignments

	Out after	Built from	One line	Due
Assignment 1 · Foundations & Optimization	L7	01-micrograd-mlp · 05-optimizers · L6–L7 reg	Extend the autodiff engine, train an MLP, race SGD/momentum/Adam, then close the train–val gap with weight decay / dropout / early stopping	+2 wks · viva after
Assignment 2 · Vision & Sequences	L15	07-cnn-shape-tour · 03-debug-ladder · 10-rnn-by-hand	Budget a CNN by hand, measure what pretraining is worth, diagnose three planted bugs with the debug ladder, and build a char-level sequence model	+2 wks · viva after
Assignment 3 · Transformers & Generative	L23	12-attention-by-hand · 13-nanogpt · 19-vae-mnist · 21-ddpm-2d	BPE by hand, attention from matmuls (verified), train nanoGPT + the temperature knob, ablate a VAE’s KL term, watch a DDPM denoise — then argue VAE vs GAN vs diffusion like an engineer	+2 wks · viva after

All three fit a free Colab GPU; Assignment 1 and most of Assignment 3’s written parts don’t even need that. Each is sized at hours, not days — if a part is eating a full day, you’re fighting an infrastructure problem, not a learning problem; come to office hours.

Grading model

Each assignment is worth 13 marks, split the same way every time:

• 8 marks — programming submission. One .ipynb that runs top to bottom and produces the figures and printouts each part asks for.
• 5 marks — individual post-deadline quiz/viva. A short one-on-one after the deadline where you explain your code, your plots, and your choices. The viva is where the marks for understanding actually live: a polished notebook you cannot explain scores near zero on those 5 marks.

There is no team project. The full schedule, deadlines, and mark breakdown are on the grading page ↗.

Grading philosophy

The assignments move from foundations, through vision and language, to generative modeling. Their purpose is not to produce a leaderboard result: each asks for a small reproducible system plus a clear account of what its measurements mean.

Within each assignment, notice where the marks actually are. The plots and accuracies are the cheap part — the largest single block is the viva, where you explain, in your own words, why momentum overshoots, why the β = 0 latent space has holes, why high temperature decays into nonsense. We would rather hear a correct mechanistic explanation of a mediocre result than see a state-of-the-art number with “it worked” underneath. Concretely:

• Accuracy targets are floors, not goals. “≥ 70% on CIFAR” means “your pipeline isn’t broken”, nothing more. No marks exist above the floor.
• Evidence over assertion. Every written or spoken claim should point at one of your own figures or printouts.
• Honest negative results score. “Adam didn’t win on my problem, and here’s the plot” is graded above a textbook answer that contradicts your own figure.
• Notebooks must run. Restart & Run All before submission; a notebook that doesn’t execute top-to-bottom is graded on what runs — and you cannot viva code that never ran, so the missing marks compound.

LLM policy

You are training to work in a field that builds these tools; pretending they don’t exist would be both unenforceable and bad preparation. So the policy is honesty-based and built around one fact: the marks that matter most live in the viva — explaining your loss curves, your latent scatter, your planted-bug symptoms, out loud, consistent with figures the model has never seen.

Allowed, no questions asked:

• Asking an LLM to explain a concept, an error message, or a piece of lecture code.
• Debugging help — “here’s my traceback, what’s wrong?” is fine (though for Assignment 2 part (c), see below).
• Boilerplate: plotting code, data-loading scaffolds, LaTeX for your derivations.

Allowed with declaration: anything beyond the above — generating solution code, drafting written answers you then edit. Each submission includes a short LLM usage note (3–5 sentences): which tools, for what, and at least one place the model was wrong or you overrode it. “No LLM use” is a fine note if true. The note is never graded for how much you used — only its absence or dishonesty costs you.

Not allowed:

• Submitting LLM-generated written analysis as your own understanding. The viva assumes the notebook’s explanations are yours; if they aren’t, the gap shows in the first question.
• For Assignment 2 part (c): pasting the sabotaged script into an LLM and asking “find the bugs”. That part grades the diagnostic process (symptom → suspect → test → fix); an oracle answer with a back-filled process is easy to spot and graded as such. Using an LLM to explain what zero_grad does after you’ve found a suspect is fine — that’s understanding, not outsourcing.
• For Assignment 3 part (a): having a model execute the BPE merges. It’s a by-hand exercise; the verification step at the end exists precisely so you can check yourself without one.

The deal, stated plainly: we design the assessed parts so that doing them honestly is the efficient path — explaining your own figure is faster than laundering a model’s explanation until it matches your plot. If you find a part where that’s not true, tell us; that’s a bug in the assignment, and we fix those. The vivas are in-person and LLM-free; if your notebook understanding is real, they are easy marks. If it isn’t, they’re where the gap shows.

Logistics

• Submit one .ipynb per assignment via the course portal, named a<N>_<rollno>.ipynb.
• The viva is scheduled shortly after each deadline; bring nothing but your submitted notebook — you’ll be asked to walk through parts of it and answer follow-ups.
• Late policy: 3 grace days across the semester, then 10%/day. Grace days move the submission deadline, not your viva slot.
• Work solo (discussion is fine; the code and writing you submit — and explain in the viva — must be your own).