Recent Developments in Deep Learning

Paris Datageeks meetup, Sept 2015

Transcript

1. Recent developments in Deep Learning Olivier Grisel - Paris Datageeks - Sept. 2015
2. Outline • Deep Learning quick recap • Recurrent Neural Networks • Attention for Machine Translation • Attention and differentiable memory for reasoning
3. Deep Learning • Neural Networks from the 90’s rebranded in 2006+ • « Neuron » is a loose inspiration (not important) • Stacked layers of differentiable modules (matrix multiplication, convolution, pooling, element-wise non linear operations…) • Can be trained via gradient descent on large data pairs of input-output examples
4. Deep Learning in the 90’s sources: LeNet5 & Stanford Deep Learning Tutorial
5. x = Input Vector h1 = Hidden Activations h2 = Hidden Activations f1(x, w1) = max(conv(x, w1), 0) y = Output Vector f3(h2, w3) = softmax(dot(h2, w3)) f2(h1, w2) = max(dot(h1, w2), 0) w1 w2 f1 f2 f3 w3
6. Recent success • 2009: state of the art acoustic model for speech recognition • 2011: state of the art road sign classification • 2012: state of the art object classification • 2013/14: end-to-end speech recognition, object detection • 2014/15: state of the art machine translation, getting closer for Natural Language Understanding in general
7. ImageNet Challenge ILSVRC2014 • 1.2 million images • 1000 classes • Last winner: GoogLeNet now at less than 5% error rate • Used in Google Photos for indexing
8. Image captioning http://cs.stanford.edu/people/karpathy/deepimagesent/
9. Why now? • More labeled data • More compute power (optimized BLAS and GPUs) • Improvements to algorithms
10. source: Alec Radford on RNNs
11. Recurrent Neural Networks
12. source: The Unreasonable Effectiveness of RNNs
13. Applications of RNNs • NLP (PoS, NER, Parsing, Sentiment Analysis) • Generative Probabilistic Language Models • Machine Translation (e.g. English to French) • Speech recognition / Speech synthesis (newer) • Biological sequence modeling (DNA, Proteins)
14. Language modeling source: The Unreasonable Effectiveness of RNNs
15. Shakespeare
16. Wikipedia markup
17. Linux source code
18. Attentional architectures for Machine Translation
19. Neural MT source: From language modeling to machine translation
20. Attentional Neural MT source: From language modeling to machine translation
21. Attention == Alignment source: Neural MT by Jointly Learning to Align and Translate
22. source: Neural MT by Jointly Learning to Align and Translate
23. source: Show, Attend and Tell
24. Differentiable memory for reasoning
25. Neural Turing Machines • Google DeepMind, October 2014 • Neural Network coupled to external memory (tape) • Analogue to a Turing Machine but differentiable • Can be used to learn to simple programs from example input / output pairs • copy, repeat copy, associative recall, • binary n-grams counts and sort
26. NTM Architecture source: Neural Turing Machines • Turing Machine: controller == FSM • Neural Turing Machine controller == RNN w/ LSTM
27. Differentiable Stack source: Inferring algorithmic patterns w Stack RNN
28. Stack RNN trained for binary addition source: Inferring algorithmic patterns w Stack RNN
29. Continuous Stack source: Learning to Transduce with Unbounded Memory
31. Reasoning for QA
32. bAbI tasks https://research.facebook.com/researchers/ 1543934539189348
33. Memory Networks source: End to end memory networks
34. source: End to end memory networks
35. source: Dynamic Memory Networks
37. Neural Reasoner
38. source: Towards Neural Network-based Reasoning
39. QA on real data
40. Paraphrases from web news
41. source: Teaching Machines to Read and Comprehend
42. source: Teaching Machines to Read and Comprehend
43. source: Teaching Machines to Read and Comprehend
44. source: Teaching Machines to Read and Comprehend
45. Conclusion • Deep Learning progress is fast paced • Many applications already in production (e.g. speech, image indexing, face recognition) • Machine Learning is now moving from pattern recognition to higher level reasoning • Generic AI is no longer a swear-word among machine learners
46. Thank you! @ogrisel