Machine Learning Research Literacy

Matrix shapes, gradients, expectations, argmin/argmax, and subscript conventions. You do not need to be able to prove things — you need to not stall on a line of algebra.

Done when: you can read a loss function aloud and say what each symbol ranges over.

MNIST or CIFAR-10, a couple of layers, written yourself rather than copied. Papers assume you have felt what a training loop, a learning rate and overfitting actually are.

Done when: it trains, you can point at the line where the gradient is applied, and you have watched it overfit on purpose.

Benchmarks, ablations, error bars (usually absent), compute budgets, and the difference between an improvement and an improvement that survives a fair baseline. Learn to check whether the comparison was tuned as carefully for the baseline as for the new method.

Done when: you can look at a results table and name two things it does not tell you.

"ImageNet Classification with Deep Convolutional Neural Networks". The result that restarted the field. Note how much of it is engineering — GPUs, augmentation, ReLU — rather than new theory.

Done when: you can explain why this paper mattered more than its accuracy number suggests, and name three of its practical tricks still in use.

Two papers about making deep networks trainable rather than more expressive. Batch norm's stated explanation has since been widely disputed — a useful early lesson that a technique can work brilliantly while its paper's account of why is wrong.

Done when: you can explain what each does mechanically, and say why batch norm's original justification is contested.

"Deep Residual Learning for Image Recognition". The skip connection, and the observation that deeper networks were getting worse for optimisation reasons rather than capacity reasons. Skip connections are now in essentially everything, including transformers.

Done when: you can draw a residual block from memory and explain what problem it solves.

One paragraph per paper, in sequence, each opening with the problem left over from the last. This is the exercise that converts four separate papers into a narrative you will still have next year.

Done when: the four paragraphs read as one continuous argument.

The default optimiser for most of a decade. Read it for the mechanism — per-parameter adaptive step sizes — and note that its original convergence proof was later found to be flawed, which did not stop it working.

Done when: you can explain what Adam maintains per parameter and why that helps.

The transformer. The most consequential ML paper of the last decade, and readable if you take it slowly. Work through the shapes by hand — attention is a small amount of linear algebra dressed in notation.

Done when: you can draw the architecture from memory and explain query, key and value in one sentence each without metaphors.

Pre-training then fine-tuning, and the idea that a general model beats a task-specific one. Note the shift in what counts as a contribution: scale and training objective rather than architecture.

Done when: you can explain masked language modelling and why bidirectionality needed a new objective.

Forty lines of PyTorch, no library attention module. Get the shapes right, apply a causal mask, and check the output against torch.nn.functional.scaled_dot_product_attention.

Done when: your implementation matches the reference to floating-point tolerance.

Pair something that argues capability grows predictably with compute against something that documents where models fail. Reading an optimistic and a sceptical paper together is the fastest way to calibrate.

Done when: you can state one claim each makes that the other would dispute.

Pass one on all three. Read only the one that survives. Deciding not to read two of them is the skill being practised — the frontier produces far more than anyone can read.

Done when: two are rejected with a written reason each, and one is read properly.

Many ML papers ship code. Clone it, run the smallest experiment, and compare against the reported number. Discovering the gap between a paper's claim and its repository is a genuine research skill.

Done when: you have a number of your own next to the paper's, and an explanation for any difference.

Summary, strengths, weaknesses, questions for the authors, and a recommendation. Real review forms are public — use one. This is how you find out whether you understood the paper or merely enjoyed it.

Done when: the review is written and every weakness is specific enough that an author could respond to it.