Sign inSign up
Beyond Seq2Seq with Augmented RNNs thumbnail
Pause
Mute
Subtitles
Playback speed
0.25
0.5
0.75
1
1.25
1.5
1.75
2
Full screen

Beyond Seq2Seq with Augmented RNNs

Published on Aug 23, 201620760 Views

Edward Grefenstette

Sequence to sequence models in their most basic form, following an encoder-decoder paradigm, compressively encode source sequence representations into a single vector representation and decode this re

Deep Learning Summer School 2016 - Montreal

Related categories

Deep LearningReinforcement LearningUnsupervised Learning

Chapter list

Beyond Sequence to Sequence with Augmented RNNs00:00
The plan02:22
The Bottleneck04:06
Some Preliminaries: RNNs - 104:06
Some Preliminaries: RNNs - 204:50
Some Obvious RNN Applications05:27
Transduction with Conditional Models - 106:40
Transduction with Conditional Models - 207:50
Sequence to Sequence Mapping with RNNs - 107:57
Sequence to Sequence Mapping with RNNs - 208:41
Sequence to Sequence Mapping with RNNs - 308:54
A Simple Encoder-Decoder Model09:34
Deep LSTMs for Translation09:45
Learning to Execute10:37
The Bottleneck for Simple RNNs11:30
Limitations of RNNs: A Computational Perspective14:22
Computational Hierarchy14:38
RNNs and Turing Machines - 116:53
RNNs and Turing Machines - 219:07
RNNs and Finite State Machines - 124:22
RNNs and Finite State Machines - 229:21
Why more than FSM?30:36
Untitled32:55
Questions? - 133:30
RNNs Revisited33:41
RNNs: More API than Model - 134:11
RNNs: More API than Model - 234:36
RNNs: More API than Model - 335:00
RNNs: More API than Model - 435:08
RNNs: More API than Model - 535:26
RNNs: More API than Model - 636:25
The Controller-Memory Split - 136:55
The Controller-Memory Split - 237:54
Attention: ROM38:25
Attention38:33
Attention (Early Fusion)39:25
RNN: X ⨉ P → Y ⨉ N40:05
Attention (Late Fusion)41:47
ROM for Encoder-Decoder Models - 142:49
RNN: X ⨉ P → Y ⨉ N44:36
ROM for Encoder-Decoder Models - 244:50
Skipping the bottleneck - 144:51
Skipping the bottleneck - 245:07
Recognizing Textual Entailment (RTE)45:41
Word-by-Word Attention46:40
Girl + Boy = Kids47:21
Large-scale Supervised Reading Comprehension48:15
Machine Reading with Attention48:58
Example QA Heatmap49:23
Attention Summary49:58
Questions? - 251:45
Untitled52:34
Untitled52:46
The Controller-Memory Split52:48
Controlling a Neural Stack53:15
Controller API53:45
Controller + Stack Interaction54:49
Example: A Continuous Stack57:55
Synthetic Transduction Tasks01:03:07
Synthetic ITG Transduction Tasks01:03:53
Rapid Convergence01:11:54
Differentiable Stacks / Queues / Etc01:12:10
Results01:13:03
Neural PDA Summary01:13:44
Register Machines: RAM01:15:36
Computational Hierarchy01:15:59
Attention as ROM01:16:10
Register Memory as RAM01:16:28
Neural RAM: General Idea01:16:44
RNN: X ⨉ P → Y ⨉ N (an example of)01:19:39
Extensions01:20:33
Relation to actual Turing Machines01:20:35
Register Machines and NLU01:23:18
Conclusions01:23:31
Thanks for listening!01:25:23