CNN-2: Progress over Training Time
Created on June 2|Last edited on April 13
Comment
Compare predictions after 1 vs 5 epochs
Top row, id 0: 1 epoch vs Bottom row, id 1: 5 epochs
These results are filtered to show only images for which the two models make different predictions, with the 1-epoch model 0 in green and the 5-epoch model 1 in violet. For example, the first row shows 5 of the 11 total images which model 0 classified as 9s but model 1 classified as 7s. Scroll right to see the full score distributions for each pairwise confusion.
Run set
2
Top errors over time
Left side: 1 epoch vs Right side: 5 epochs
See incorrect predictions (filter to rows where "guess" != "truth") on the full test data. Note that there are 229 wrong guesses after 1 training epoch, but only 98 after 5 epochs.
Run set
2
False positives grouped by guess
Left: double layer size & learning rate; Right: baseline
Filter out correct answers, then group by the guess to see examples of misclassified images and the underlying distribution of true labels—for two models side-by-side. A model variant with 2X the layer sizes and learning rate is on the left, and the baseline is on the right. Note that the baseline makes slightly more mistakes for each guessed class. Scroll to the right in each of the two vertical panels to see the score distributions for other classes.
Run set
2
Add a comment