The term project of Introduction to Machine Learning asks each team to adopt one published machine learning paper and spend the semester with it. The team reads the paper until it can retell it in its own words, rebuilds the experiments to see how faithful such a rebuild can be, questions the theory where it is thin, and demonstrates the found weaknesses with experiments of its own design. The work runs as three milestones and two labs, each with its own page, and the course closes it at the final oral exam.
The final oral exam¶
The route¶
The checkpoints run in the order below, each with a dedicated page. The milestones set up the knowledge, and the labs turn it into running experiments:
Selecting the Paper: the team registers and claims its paper by submitting the selection form published on Moodle.
Reading the Paper: the team internalizes the paper and hands the retelling in as a slide deck.
Experimental Reproduction: the team reruns the paper’s experiments and measures how close its rebuild lands to the published numbers.
Critical Analysis: the team critiques the theory of the paper, from its assumptions to its complexity story.
Empirical Demonstration: the team demonstrates the found weaknesses empirically and compresses the whole critique into one poster.
The checkpoints pair up by deliverable: the reading milestone and the reproduction lab submit together as one slide deck, and the critique milestone and the empirical-demonstration lab submit together as one poster.
Logistics¶
The project is a team effort. The selection form captures every member, and the confirmed paper binds its team for the rest of the semester.
Each team keeps one public code repository for the entire semester. Code, notebooks, and documentation accumulate in it from the first week, and the commit history documents the progress. Everything the deck and the poster report must trace back to this repository.
Submissions leave through Moodle: the selection form at the first milestone, the slide deck file and the repository link at the deck checkpoint, and the poster file at the poster checkpoint. Deadlines follow the Moodle announcements of the course.
Experiments are tracked run by run on a public Weights & Biases project so that every reported number in the deck or the poster links to its run. The tracking rules live in Experimental Reproduction and apply to both labs.
Grading¶
| Part | Weight |
|---|---|
| The selection form | pass gate, no points of its own |
| The slide deck, the reading milestone and the reproduction lab together | 45% |
| The poster, the critique milestone and the empirical-demonstration lab together | 50% |
| Tracked repository quality, W&B discipline, and artifact formatting | 5% |
| Total | 100% |
The selection form carries no points of its own and gates the semester: a team that has not registered its paper earns no grade for the project. The final rubric row scores the working discipline around the artifacts rather than the artifacts themselves, so keep the repository readable, the experiments tracked, and the formatting of both artifacts tight.
An excellent deck retells the paper precisely in the team’s own words, visualizes the method faithfully, and pins the reproduction results against the published ones with an account of every deviation, to the point that the examiners could reconstruct what the paper does without opening it. A deck that retells correctly but visualizes thinly, or reports deviations without explaining them, reads as unfinished work. An excellent poster puts a real weakness on the table, argues it theoretically, and confirms it with an experiment designed exactly for that weakness. A poster that asserts weaknesses without theory, or shows experiments disconnected from any claimed weakness, reduces the semester to unbacked opinions.
Integrity and the use of AI¶
Reimplement the proposed method from the paper’s description. If the paper ships a public implementation, that code serves verification and comparison only. Copying the paper’s text or its figures, copying code from the original implementation, or copying the work of another team zeros the project.
As usual, AI tools are not restricted: use them wisely. They may support the writing and the coding, but every member stands behind the whole artifact at the poster workshop and explains any part of the work there. The lab instructor reserves the right to conduct additional assessments on some of the teams, should the work of a team make them needed.
Each checkpoint page specifies its own deliverable: the selection form in Selecting the Paper, the slide deck in Reading the Paper, and the poster in Empirical Demonstration.