/margin-bounds

Primary LanguageJupyter Notebook

A visualization of a statement in "Spectrally-normalized margin bounds for neural networks".

In section 2, they mention a useful metric for the difficulty of a dataset,

The distribution of this over the training set somehow captures the difficulty of a dataset.

This is a visualization of that quantity with MNIST and FashionMNIST. I only change which dataset is being used.