Spotlight
When does label smoothing help?
Rafael Müller · Simon Kornblith · Geoffrey E Hinton

Thu Dec 12th 10:30 -- 10:35 AM @ West Ballrooms A + B

The generalization and learning speed of a multi-class neural network can often be significantly improved by using soft targets that are a weighted average of the hard targets and the uniform distribution over labels. Smoothing the labels in this way prevents the network from becoming over-confident and label smoothing has been used in many state-of-the-art models, including image classification, language translation and speech recognition. Despite its widespread use, label smoothing is still poorly understood. Here we show empirically that in addition to improving generalization, label smoothing improves model calibration which can significantly improve beam search. However, we also observe that if a teacher network is trained with label smoothing, knowledge distillation into a student network is much less effective. To explain these observations, we visualize how label smoothing changes the representations learned by the penultimate layer of the network. We show that label smoothing encourages the representations of training examples from the same class to group in tight clusters. This results in loss of information in the logits about resemblances between instances of different classes, which is necessary for distillation, but does not hurt generalization or calibration of the model's predictions.

Author Information

Rafael Müller (Google Brain)
Simon Kornblith (Google Brain)
Geoffrey E Hinton (Google & University of Toronto)

Geoffrey Hinton received his PhD in Artificial Intelligence from Edinburgh in 1978 and spent five years as a faculty member at Carnegie-Mellon where he pioneered back-propagation, Boltzmann machines and distributed representations of words. In 1987 he became a fellow of the Canadian Institute for Advanced Research and moved to the University of Toronto. In 1998 he founded the Gatsby Computational Neuroscience Unit at University College London, returning to the University of Toronto in 2001. His group at the University of Toronto then used deep learning to change the way speech recognition and object recognition are done. He currently splits his time between the University of Toronto and Google. In 2010 he received the NSERC Herzberg Gold Medal, Canada's top award in Science and Engineering.

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors