Show how training error decreases over time

[JoshRosen]

We should add a view that shows how the training error decreases over time. Some ideas:

• Add a sparkline showing the training loss versus time. This could be drawn parallel to the timeline slider.
• Add a confusion matrix view that changes in response to the timeline slider.

[JoshRosen]

Implementing the confusion matrix could be difficult because it looks like cuda-convnet's GPU workers only return the per-batch loss rather than the actual misclassifications.

[JoshRosen]

I implemented a confusion matrix in e6eb6e846cc9192ab81e19eee724e7af5d1204bc. I'll upload a snapshot of the database to Dropbox in a little bit so you can test it out. I'd still like to add mouseover support for viewing a sample of the images that fall into each misclassification category (for example, mouseover to view 9 cats that were misclassified as dogs).

[JoshRosen]

f118f4d0692b5919bbe303a243662ce00f9709eb added support for viewing a sample of misclassified images.

Here's a link to a recent build of the stats_db: https://www.dropbox.com/s/t5zg9dqb0eoe6lv/stats_db.zip

Unizp that file and use that directory as the --model-stats parameter for runserver.py if you want to load the data into the app.

[JoshRosen]

I think there's a bug in the confusion matrix's image display because the images' classes displayed in the histograms don't match up with the classifications reported by the table. I'll investigate this afternoon.