Skip to main content

Assignment 2: Break the number

Construct a bench candidate that scores well and is visibly wrong, or scores badly and is visibly fine. Submit the set, the score and the mechanism.

Due · Week 9

Course weight

25%

What to submit

Submit one PDF of at most three pages, and one archive that rebuilds your candidate from a single seed on a clean machine.

Build something the score gets wrong, then show exactly how you did it.

A candidate that scores well and is visibly wrong, or one that scores badly and is visibly fine. Either direction counts. What is marked is whether you can name the part of the machinery you exploited, and whether the object you built matches the account you give of it.

Due end of week 9. Deadlines, extensions and the use of generative tools are course-wide and live on the policies page.

Before you start

  1. Individual work. 100 points, worth 25 per cent of the course.
  2. Three parts. Part A is 40 points, Part B 40, Part C 20. The marking criteria on this page are those three parts.
  3. The parts are not independent. Part B must explain the object built in Part A, and Part C must report on both. A mechanism that does not match the construction earns nothing for either.
  4. Generative tools are governed by the policies page. Read it before you start rather than before you submit.
  5. Submit one PDF of at most three pages, and one archive that rebuilds your candidate from a single seed on a clean machine.

Part A, the construction [40]

Build a bench candidate that scores well and is visibly wrong, or scores badly and is visibly fine. “Visibly” means by a per-sample measure that you define in this part and defend.

Score it against the bench N-honestly, with both FID and FID∞, and state N and the implementation for each number.

Candidate C from week 6 is the worked example. It is not an acceptable submission.

Part B, the mechanism [40]

Say which layer your construction breaks.

If it is the estimator, show the bias term and how your candidate inflates it. If it is the quantity, show which moments match and what the score therefore cannot see. Name the week that predicted it.

One layer argued properly beats two named loosely. The stronger submissions tend to break the quantity, because a broken estimator can be repaired by measuring more and a broken quantity cannot be repaired at all.

Part C, the write-up [20]

What would a reader of a results table have needed to see in order to catch your candidate? One page.

This is the part that becomes your final report.

In person

Ten minutes, week 10, at the codewalk, which carries the day, the rooms and the booking. Your archive is run from its seed on our machine and the score must reproduce to one decimal place. Then one question on Part B. Bring your ID.

Supporting experiment / Workspace

A perfect score cannot certify originality.

  1. 01 Predict

    If the candidate copies the evaluated reference collection, what happens to empirical FID? What will shuffling that collection do to FID and the paired distance?

  2. 02 Open the lab

    Image lab / identity and shuffle

    Use the supplied identity control and measure once. Shuffle candidate order, then restore it. No new feature extraction is needed for the shuffle.

  3. 03 Record

    Export the counts, identity mode, feature vectors and candidate ordering before and after shuffling. Keep a note of the source images; the export does not contain the original images.

  4. 04 Result

    Compare empirical FID with the separately labelled mean paired squared feature distance. Identify which quantity depends on displayed pairing.

  5. 05 Explain the limitation

    Exact copying matches empirical moments; independent samples or an arbitrary subsample need not. A near-zero numerical score does not certify originality. Shuffling alone is not an A2 counterexample, and this is not the mixture C population construction.

Keep this supporting experiment distinct from the task above. Workspace use and reporting guidance →

Submission review

Spec

The brief poses the problem and leaves room for a response. The spec is the fixed part: what a marker checks. Some lines can be checked mechanically; the rest call for judgement.

Tick each requirement after reviewing your evidence. These ticks record your review, not a grade or confirmation that the requirement is met.

Without JavaScript, ticks last until you leave this page.

How it is marked

CriterionWeight
The construction40%
The mechanism, stated in terms of the estimator or the quantity40%
The write-up20%