Timezone: »
The recently proposed Temporal Ensembling has achieved state-of-the-art results in several semi-supervised learning benchmarks. It maintains an exponential moving average of label predictions on each training example, and penalizes predictions that are inconsistent with this target. However, because the targets change only once per epoch, Temporal Ensembling becomes unwieldy when learning large datasets. To overcome this problem, we propose Mean Teacher, a method that averages model weights instead of label predictions. As an additional benefit, Mean Teacher improves test accuracy and enables training with fewer labels than Temporal Ensembling. Mean Teacher achieves error rate 4.35\% on SVHN with 250 labels, better than Temporal Ensembling does with 1000 labels.
Author Information
Antti Tarvainen (The Curious AI Company)
Harri Valpola (The Curious AI Company)
Related Events (a corresponding poster, oral, or spotlight)
-
2017 Poster: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results »
Wed. Dec 6th 02:30 -- 06:30 AM Room Pacific Ballroom #13
More from the Same Authors
-
2017 Poster: Recurrent Ladder Networks »
Isabeau PrĂ©mont-Schwarz · Alexander Ilin · Tele Hao · Antti Rasmus · Rinu Boney · Harri Valpola