Rethinking Generalization: Shuffling the Labels
Beginner
The problem, from first principles
Zhang et al. (2017) showed that neural networks can fit completely random labels perfectly, memorizing the training set. Shuffling the labels is the first step of that experiment: the labels keep their distribution but lose their connection to the inputs.
From theory to code
Implement shuffle_labels(y, rng), which returns a random permutation of the labels.
Constraints
- Use the provided
rng, so results are reproducible.
Hints
Hint 1
Call rng.permutation on a NumPy copy of the labels.
Keywords
research-papersclassical-mllearning-theorygeneralization
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.