Homework 1 ✍🏻

Back to the AGI Course - 2026 Fall (AI5102)

The first homework asks you to examine whether the problem you are working on is defined the way you actually want it defined. It is due on Friday, Oct 2, at 11:59 PM.

  • Deliverables:
    • Write your answer by hand on A4 paper, 2-4 pages (longer is not allowed). Scan it or take photos, and submit a single PDF to the Gradescope entitled “HW1”. Typed submissions will get significant deduction.
    • Write your name, student ID at the top.
    • Notation, equations, and hand-drawn diagrams are all welcome. You may write formally or in prose.
    • Write in English.
    • You may use an AI assistant while preparing your answer. If you do, add a paragraph at the end describing how you used it.
    • See this video for a Gradescope tutorial. If there are multiple questions, match each page to the corresponding question when you submit.

  • Purpose of this assignment:

    In the lectures so far, you have heard how the instructor narrowed and reshaped a problem over the course of a research project. What repeatedly mattered was not the choice of method but the decision of how far to narrow the problem. The expensive mistake in research is not picking a bad method. It is inheriting someone else’s narrowed problem without knowing why it was narrowed. A benchmark existing means someone already traded away several axes, and whether they did so because those axes are inessential or because they were infeasible at the time. So, undo one of those trades, write down the resulting problem, and see how much of your current approach survives it.

  • Contents:

    Track A — if you already have a research topic.

    1. The problem as it stands: State the problem you are actually working on. Not the framing you would put in a paper introduction, but the question this week’s experiments actually answer. Then list the axes that have already been traded away — e.g., the form of the input (single image vs. video stream), the number of targets (single vs. multi-object), offline vs. online operation, the evaluation metric, assumptions about the data distribution, compute or latency budget, the degree of human involvement.
    2. The problem with a trade undone: Redefine the problem with at least one of those axes restored. Set feasibility aside for now. Be explicit about inputs, outputs, constraints, and what counts as success. Formal notation is welcome.
    3. The gap: Under this new definition, how effective is the approach you have been developing? Does it extend, or does it break once a particular assumption is removed? Name where it breaks, concretely — not “this would probably be hard,” but something like “this method assumes labels are independent across frames, which does not hold in a stream.”
    4. What to think about next. Pick one gap from (3) worth attacking first, and say why that one. This is not a request to abandon what you are doing. It is a request to locate your current work within the larger picture.

    Track B — if you do not have a topic yet.

    1. If you could work on one problem freely, what would it be? Write its definition at the level asked for in A-2.
    2. Find a smaller version of it that people around you, or the community, are working on — a paper, a project in a lab that you are interested in. Describe it.
    3. Where exactly do the two differ? Which axis did they trade away, and was the trade reasonable? If it was reasonable, why do you still think the wider definition is the right one?
    4. What would be your first step toward your version of the problem?

    Any problem is fine as long as it is not completely untethered. It does not have to be an AI/ML problem.

  • An example:

    To be useful in practice, this has to run as multi-object instance segmentation on a video stream. But the community is improving single-object classification + localization, so I have also been testing combinations of method A and method B for marginal gains on a known model. The axes traded away are (i) the number of objects, (ii) the time dimension, and (iii) identity across frames. Restoring (ii) —

  • Answers that will not score well:
    • The definition gets grander in wording but not wider in scope (“build an agent with general intelligence”). A wide definition should still be able to say what goes in and what comes out.
    • (1) and (2) describe the same problem, meaning no axis was actually restored.
    • (3) concludes that your method would work fine under the new definition — in which case the definition was not widened enough.
    • Most of the pages are related work, background, or survey. This assignment is about your judgment.
  • Grading Criteria (15%):
    • Check Plus (13-15%) - The traded axes are identified precisely, the extended definition is concrete, and the limits of the method are stated honestly.
    • Check (10-12%) - Adequate. Most submissions are expected to fall into this category.
    • Check Minus (7-9%) - The definition is vague, the extension is nominal, or the diagnosis is superficial, or the writing is too verbose.
    • Typed submission (3-6%) - 60% deduction if your submission is typed.
    • No submission (0%)
    • Late submission will be graded according to the late policy