Reading the Paper

Objectives¶

Reading a paper is more than parsing its sentences. By the end of this milestone, the team holds a working understanding it can defend without the PDF open: what problem the paper attacks, what gap it fills, what is genuinely new in its method, how the method works part by part, and how the authors evaluate it. This milestone shares its deliverable with Experimental Reproduction: the two submit together as one slide deck, where the reading supplies the retelling and the lab supplies the running experiments.

The reading guide¶

Work through the paper in the order below. Every element gets retold in the team’s own words, and the retelling must survive in front of an examiner who has not read the paper. Verbatim sentences out of the paper do not count as an answer. The point of the milestone is the conversion.

A fast pass before the close reading¶

Read the paper in the recommended order: the abstract, then the introduction, then the conclusion, then the background and related work, then the experiments, and the method last. The first three tell you what the paper promises and where it lands, and that is enough to judge early whether the paper is worth the rest of the semester. The background situates the claim among its neighbors, the experiments show whether the evidence holds up before you pay for the details, and the method, the heaviest part, reads fastest once you already know what to look for. A team that cannot say what the paper promises and where it lands after this fast pass should go back to Selecting the Paper and reconsider.

Context and motivation¶

• The problem the paper attacks, stated plainly, with a concrete example of your own that shows why the problem matters in the real world.

• At least two predecessor approaches the paper builds on. For each one, state the core idea, the limitation that remains, and the practical consequence of that limitation. A limitation without its consequence is an afterthought, not an analysis.

Contributions¶

List the contributions in the paper’s own terms, then reorganize them by kind: theoretical contributions, technical contributions, and empirical contributions. Novelty is judged against the predecessors above: state for each contribution what exactly was not possible or not attempted before.

The method¶

This is the technical heart of the deck. The retelling covers five layers:

• Notation and the formal problem. Restate every symbol the paper uses in one consistent notation of the team’s own, and restate the problem the method solves, formally where the paper is formal.

• The overall pipeline. Draw the method end to end in a diagram of the team’s own making and walk the reader through the data from input to output.

• The components, one by one. Each component answers three questions: what it does, why the method needs it, and how it works precisely. Key formulas appear with every variable explained, and nothing on a slide floats without its meaning in the speaker notes.

• The theoretical properties. Restate every lemma, theorem, or proof sketch the paper claims and explain what it means in intuitive terms. A full reproduction of the proofs is not required. A working understanding of what the properties promise and where their limits sit is.

• Complexity. The time and space story of the method, and how it compares with the predecessors named above.

The combined deliverable: the slide deck¶

The reading milestone and Experimental Reproduction submit as one deck.

• One pptx file of at least 20 slides and at most 40 slides, counting the main story and any appendix slides alike. Fewer slides usually mean the team skipped the analysis. More slides usually mean the team copied the paper section by section.

• The deck is a visual artifact. It re-describes the paper visually: contributions, novelty, the gap it fills, the method, and the methodology of datasets, metrics, and protocol, each rendered as drawn figures, diagrams, and tables rather than paragraphs.

• Every slide carries speaker notes of at most 500 words. The notes hold the verbal analysis, the reasoning, and the connective tissue that the visuals cannot fit. Notes are read as submitted. There is no oral presentation, so the deck plus its notes are the complete artifact.

• Citations render in IEEE style, numbered, at the bottom of each slide that cites them. There is no separate references slide: a claim stands next to its source on the same page.

• All figures belong to the team. Redraw or recreate whatever the paper shows. Pasting the paper’s figures counts as copying under the integrity rules.

• The deck closes the story with the reproduction: the team’s results placed against the paper’s, each team number citing its W&B run, as told by Experimental Reproduction.

How the deck gets read¶

The examiners open the deck, read it slide by slide against the speaker notes, and check the retelling against the paper where the retelling makes claims. A deck that restates the paper earns little. A deck that shows the team understood, drew it, and tested it earns the milestone. Slides that dump raw source code fail the visual requirement, and slides without speaker notes count as incomplete.