arXiv Analytics

Sign in

arXiv:1712.05652 [cs.LG]AbstractReferencesReviewsResources

Pre-training Attention Mechanisms

Jack Lindsey

Published 2017-12-15Version 1

Recurrent neural networks with differentiable attention mechanisms have had success in generative and classification tasks. We show that the classification performance of such models can be enhanced by guiding a randomly initialized model to attend to salient regions of the input in early training iterations. We further show that, if explicit heuristics for guidance are unavailable, a model that is pretrained on an unsupervised reconstruction task can discover good attention policies without supervision. We demonstrate that increased efficiency of the attention mechanism itself contributes to these performance improvements. Based on these insights, we introduce bootstrapped glimpse mimicking, a simple, theoretically task-general method of more effectively training attention models. Our work draws inspiration from and parallels results on human learning of attention.

Comments: Presented at NIPS 2017 Workshop on Cognitively Informed Artificial Intelligence
Categories: cs.LG, cs.CV
Related articles: Most relevant | Search more
arXiv:1804.01653 [cs.LG] (Published 2018-04-05, updated 2018-08-28)
Review of Deep Learning
arXiv:1710.06319 [cs.LG] (Published 2017-10-17)
Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks
arXiv:1611.10176 [cs.LG] (Published 2016-11-30)
Effective Quantization Methods for Recurrent Neural Networks