ML wingM07
Machine learning since 2023. Every result shared in the wing says which data it was tested on.
How the model decides
Forty-one neighbours vote on every point, so the regions are two smooth bands. This is the worst model here: too simple to follow the real shape of the data, even though it looks the most sensible.
With fewer voters, the boundary starts to follow the data. Both scores improve together, and it is tempting to think they always will.
Seven neighbours: the best score on data the model has not seen. You can only find this point if you kept some data aside before you started.
Training accuracy keeps rising, but accuracy on unseen data is already falling. Only the second number tells you how the model will do in real use.
Three neighbours, and small islands appear around single points. The model is now fitting this particular sample instead of the two groups.
One neighbour. The model scores 100% on its training data, because every point is its own nearest neighbour, but on unseen data it is back to where the simplest model started.
The matrix · kNN · k = 7 · held-out split Example data
Actual class
cat
40 of 48correct
Mistaken for: dog 8
Two-thirds of the errors come from one pair the model cannot tell apart: cats and dogs. An accuracy figure would never show which pair, and that pair is what to work on next.
Events
Best known for
Build a full AI app, not just a notebook: train a model, then put it inside working software. Each team has exactly one ML specialist and recruits the rest from the blockchain or web tracks. Topics range from deep learning to GenAI, multimodal and NLP. Copied repositories were disqualified.
A production AI system handed over broken. It has four parts, an ML pipeline, a RAG chatbot, a FastAPI backend and a Streamlit interface, all sharing one config file. Each part has hidden bugs. Fix it before time runs out. Open to first-years.
You get a topic and a persona, and use your own LLM setup to argue for or against the topic. Entries are sent as structured JSON. A critic model ranks them on logic, clarity, creativity and whether the persona stayed consistent. Open to first-years.
A deep learning contest on Kaggle, ranked by leaderboard accuracy. You build CNNs from scratch, with no pretrained weights and no transfer learning. You get ten submissions a day, and your best two count. Open to first-years.
Files
No class files yet.
Class log
Team
2 coordinators · 4 senior members · 5 members
Coordinator
No earned number yet
Coordinator
Google Summer of Code HumanAI
Senior member
No earned number yet
Senior member
Research Fellow SPAR
Senior member
No earned number yet
Senior member
No earned number yet