/gen-viz

Code for the paper "Understanding Generalization through Visualizations"

Primary LanguagePython

Visualizing Generalization Phenomena

The code for this paper is split up into different directories for each section of the paper, and each directory has its own README for how to run the code.

  • minefield contains the code to train a good model and look for bad minima along its trajectory (Figure 1)
  • linear_model contains the code for training the overparametrized linear model and neural net (Figure 2)
  • optimizers contains the code for training on CIFAR using different optimizers (Figure 2, Table 1)
  • swissroll contains the code for all figures on the swiss roll dataset (Figures 3, 5, 6)
  • svhn contains the code for all figures on the SVHN dataset (Figures 4, 5, 7)
  • concentric_circles contains the code for the concentric circles wide margins experiments (Figure 8)