Run a Backpropagation Bug-Hunt on an Open-Source RL Implementation
Overview
What this challenge is about.
You receive the offending fork (around 4,000 lines of PyTorch) and three known-failure seeds. Reproduce the NaN failure deterministically, instrument the forward and backward pass per layer (PyTorch hooks), identify which custom layer's gradient is wrong, propose and verify a fix, and write a 4-page post-mortem with a CI plan that would have caught the bug.
The Brief
What you'll do, and what you'll demonstrate.
Identify and fix a backpropagation bug in a custom layer of an open-source RL library and propose CI that prevents recurrence.
Earning criteria — what you'll demonstrate
- Debug numerical instabilities in deep-learning training
- Instrument forward and backward passes with PyTorch hooks
- Design unit tests that catch backpropagation bugs
- Write engineering post-mortems that drive systemic improvements
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.
Research Scientist
Backpropagation bug-hunts and rigorous post-mortems are exactly the kind of work research scientists do to harden lab infrastructure.
This challenge sharpens
- backpropagation
- pytorch
- debugging
ML Researcher
Numerical-stability debugging is increasingly required for any ML researcher running long training jobs at scale.
This challenge sharpens
- backpropagation
- numerical-stability
- debugging
MLOps Engineer
Designing CI that catches gradient bugs is exactly the kind of platform improvement MLOps engineers ship.
This challenge sharpens
- ci-design
- post-mortem-writing
- pytorch