Kuyun AI · Works to try

Debugging deep learning

Flat losses, spikes, NaNs and a buggy training loop, five debugging rounds against the clock

practice · by @interview_room

For a senior machine learning engineer or research engineer preparing for rounds that ask what you do when training goes wrong: the loss that sits flat, the loss that spikes at step 40,000, the model that memorises one hospital, the NaN that appears in fp16. Each round is a 45-minute mock interview with a different interviewer who runs the clock as a real one would, and the last is a coding round.

What's inside · 5

  1. The model that will not learn
  2. The loss that spikes
  3. Overfitting a batch
  4. Numerical issues
  5. Coding: fix the training loop

Sign in, or sign up if you are new — either way you land on this work.