Member-only story
How we beat the FastAI leaderboard score by +19.77%…a synergy of new deep learning techniques for your consideration.
After months of testing various new deep learning activations, optimizers, and more, our team combined multiple cutting edge techniques to soundly beat the long standing FastAI 5 epoch ImageWoof leaderboard score.
(Update — we have now set high scores for a total of 12 categories on the FastAI leaderboard, for both ImageWoof and ImageNette datasets. One of our techniques used, Flat+Cosine Anneal, is also now officially in the FastAI codebase).
This article details the new deep learning techniques we used to accomplish this, and introduce them to you for your consideration in your own work.
The previous record for 5 epochs was 55.2%. We achieved 74.97%, or a jump in accuracy of +19.77%. The leader-board rules require at least 1% improvement to enter a new score, and even that can be hard to do…so we were extremely pleased with our final results.
