Problem Set 2

The pre-proposal happens here

Problem 6 of PS2 is the first piece of your final project, and it is sharpened here in conversation before you post it. The other problems are on intro2ml.com and go to Moodle as usual.

Problem 6: Project Pre-Proposal

10 points. About 300 words, five things, written before you open this conversation. Read the project guidelines first.

The five things
  • Title and description. The idea in a few sentences
  • Problem. What you would predict or model, and why it is interesting
  • Data. A dataset you could actually get, with a link if you have one
  • Hypothesis or question. What you would test, and one or two references or similar projects you build on
  • Plan. The first method you would try, and how you would know it worked
Where drafts usually fail
  • A topic instead of a question: interesting area, nothing that could come out false
  • Data named vaguely ("social media data") with no path to actually getting it
  • A first method from the end of the course, with no baseline from the start of it
  • No idea what success would look like: which metric, on which held-out data, compared with what

Finding a dataset is usually easier than finding an idea. If you are stuck, start from data you can actually get and let the question grow from it. A modest problem on real data beats an ambitious one nobody could finish in six weeks.

Write it first, then paste it

  • The draft is yours. Write the page before you open the conversation. The agent will not write it for you, and asking it to will cost you the marks this problem is made of.
  • It will push back. On whether you have a question or only a topic, whether the data exists and you can get it, and whether your first method could answer the question.
  • Argue. It is often wrong about your problem. Say so, and hold the position you can defend.
  • Paste revisions, do not describe them. When you rewrite a section, paste the new words in. It reads the recent conversation, not an archive of it.
  • It can search, and it will. When a claim about a dataset or a reference matters, it looks it up and tells you what it found. It will still ask what the data has in it and what the baseline is, which is the part only you can answer.
  • Then post. The revised pre-proposal goes on #iml-project, ending with one or two sentences on what the exchange changed your mind about.

What a good exchange looks like

The mark is not for a smooth proposal. It is for a proposal that survived being argued with, and for knowing which parts did not survive.

  • Defend the question, not the topic. If you cannot say what result would make you abandon the hypothesis, you have a topic.
  • Bring the dataset, not the impression. When it asks what the data has in it, answer from the data: rows, columns, where it comes from.
  • Let the scope shrink. Most first drafts are three projects. The useful outcome is often one of them, done properly.
  • Keep what you disagreed with. Those sentences are the ones worth writing up at the end of your post.

Then it goes to Slack

Post the revised pre-proposal on #iml-project. Read your classmates' posts while you are there and comment on at least two: the projects that go well are usually the ones that found a neighbour early, and the teams you form next start from these posts.

Found a bug or a badly worded question? Use the comment button at the top right. It goes straight to a list I read.