Overview
What this challenge is about.
Pick one of three pre-approved 2025 papers (offered by the supervisor) with a known reference codebase you may consult but not copy. Re-implement the model and training loop in PyTorch from the paper alone for the first 5 days, then permit yourself to peek at the reference repo to debug. Train on a public dataset subset that fits in a single L4 GPU-day. Report top-1 accuracy on the standard validation split alongside the paper's reported number, and write a 3-page reproducibility note covering every deviation, surprise, and gotcha.
The Brief
What you'll do, and what you'll demonstrate.
Re-implement a recent SOTA vision paper from scratch in PyTorch, report the reproduction gap, and document every deviation honestly.
Earning criteria — what you'll demonstrate
- Read an ML paper closely enough to implement it
- Debug training failures (loss not decreasing, gradient instability) in PyTorch
- Quantify the gap between reported and reproduced numbers honestly
- Communicate reproducibility caveats to a technical audience
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
ML Researcher
Reproducing a paper end-to-end and honestly documenting the gap is the rite-of-passage exercise every junior ML researcher is expected to have done at least once.
This challenge sharpens
- paper-reproduction
- pytorch
- scientific-writing
Research Scientist
The discipline of training-debugging and explicit deviation logging mirrors the daily rigor of a research scientist running ablation studies.
This challenge sharpens
- experiment-design
- training-debugging
- scientific-writing
Applied AI Scientist
Translating a paper into running code on a constrained budget is exactly the work applied AI scientists do when bringing fresh research into a product team.
This challenge sharpens
- pytorch
- deep-learning
- paper-reproduction