BlackOut: Speeding up Recurrent Neural Network Language Models With Very Large Vocabularies thumbnail
Pause
Mute
Subtitles not available
Playback speed
0.25
0.5
0.75
1
1.25
1.5
1.75
2
Full screen

BlackOut: Speeding up Recurrent Neural Network Language Models With Very Large Vocabularies

Published on May 27, 20162117 Views

We propose BlackOut, an approximation algorithm to efficiently train massive recurrent neural network language models (RNNLMs) with million word vocabularies. BlackOut is motivated by using a discrimi

Related categories

Chapter list

BlackOut: Speeding up RNNLMs with Very Large Vocabularies00:00
Prevalence of Large Softmax Output Layers00:22
Case Study: RNNLM01:35
System Optimization 04:50
Strategies to Speed up Softmax05:55
Blackout Training06:51
Connection to Importance Sampling09:28
Connection to Noise Constrastive Estimate (NCE)10:45
Comparison to Dropout12:04
Experiments on Small Datasets - 113:34
Experiments on Small Datasets - 214:34
Experiments on 1-Billion Word Benchmark14:41
Comparison to State-of-The-Arts15:15
Conclusion16:33