Kaggle Computer Vision Research
Neural Debris Removal Competition
A private, in-progress competition workspace for removing learned artifacts from a provided object detector. The project combines model reconstruction, controlled unlearning experiments, held-out evaluation, object-level routing, submission validation, and reproducibility checks. Detailed strategy remains private while the competition is active.
Period
Jul 2026 - Present
Tools / Tech
PythonPyTorchTorchvisionRetinaNetScikit-learnNumPyPandasPytest
Why I built it
I treated the task as an experimental-design problem: first reproduce the supplied model closely enough to trust local comparisons, then test one change at a time and use limited leaderboard feedback only as secondary evidence.
What it includes
- Reconstructs a RetinaNet-compatible inference path and checks local predictions against supplied reference outputs before accepting training results.
- Organizes full-model, classification-head, distillation, pseudo-clean, and routing experiments through versioned configurations and held-out comparisons.
- Validates submission schema, freezes selected artifacts with hashes, and records both successful and rejected directions in a decision log.
- Keeps raw competition data unchanged and separates verified competition facts from assumptions that still require evidence.
What I worked on
- Defined the experimental gates, retention-versus-removal tradeoffs, comparison criteria, and the order in which candidate ideas were tested.
- Used AI assistance to accelerate implementation and experiment plumbing, while I reviewed the model behavior, chose follow-up experiments, interpreted failures, and decided which candidates were worth submitting.
- Built the workflow around reproducible evidence: cross-fitted checks, deterministic submission generation, automated tests, provenance manifests, and explicit rollback points.
- Tracked a substantial public-leaderboard improvement from the first submitted baseline and kept the project marked as active work.
What I Learned
- Learned why a faithful local baseline is necessary before interpreting small experimental improvements.
- Practiced changing direction when leaderboard evidence contradicted a local proxy instead of defending the original idea.
- Built stronger habits around experiment provenance, hidden-label uncertainty, and communicating incomplete competition results carefully.