demystifying deep learning - roberto paredes palacios @ papis connect

24
Deep Learning Roberto Paredes ([email protected]) PRHLT Research Center Universitat Polit` ecnica de Val` encia March 2016

Upload: papisio

Post on 16-Apr-2017

410 views

Category:

Technology


1 download

TRANSCRIPT

Page 1: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning

Roberto Paredes ([email protected])PRHLT Research Center

Universitat Politecnica de Valencia

March 2016

Page 2: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Introduction

• Neural networks

• Deep Learning: Stack many layers to build deep models

• Recently, grab the attention of the industry

• Many new applications

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 2

Page 3: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Introduction

http://qz.com/335768/bill-gates-joins-elon-musk-and-stephen-hawking-in-saying-artificial-intelligence-is-scary/

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 3

Page 4: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Introduction

• Key issue: Representational Learning

• Seamless Representation-Classification model

... and make it happen!

... and even make it affordable!

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 4

Page 5: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Introduction

• Neural Network:

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 5

Page 6: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Introduction

• Deep Learning → Bridge the gap between raw representation and categories

http://www.clarifai.com/static/img_ours/cnn.png

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 6

Page 7: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Applications

• Deep Learning:

– Key issue Representational Learning– Some realistic problems require a deep structure to be learned properly

• Applications:

– Image Recognition– Speech Recognition– Natural Language Processing– Machine Translation– ...

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 7

Page 8: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Computer Vision

• ImageNet Challenge

http://blogs.nvidia.com/blog/2014/09/18/gpus-imagenet-deep-learning/

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 8

Page 9: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Computer Vision and NLP

http://googleresearch.blogspot.com.es/2014/11/a-picture-is-worth-thousand-coherent.html

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 9

Page 10: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Computer Vision and NLP

http://googleresearch.blogspot.com.es/2014/11/a-picture-is-worth-thousand-coherent.html

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 10

Page 11: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Computer Vision

Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional

neural networks. In Advances in neural information processing systems (pp. 1097-1105).

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 11

Page 12: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Computer Vision

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 12

Page 13: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Computer Vision

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 13

Page 14: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Speech Recognition

• Tandem DBN-DNN-HMM

https://www.cs.toronto.edu/~hinton/absps/DNN-2012-proof.pdf

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 14

Page 15: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning Achievements - Handwritten Text Recognition

• Bidirectional LSTM. CTC and LM

• ICDAR competition, WER: from 27 (Basic HMM) down to 15 (Combinations)

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 15

Page 16: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning (Neural Networks) Approaches

• Neural Network (Multi-Layer Perceptron)

– Deep Neural Network (Multi-Layer Perceptron with more layers)– Word2Vec (NLP) (Multi-Layer Perceptron)– Autoencoders and Denoising Autoencoders (Multi-Layer Perceptron)

• Convolutional Neural Networks

• Long-Short Term Memory LSTM

• Restricted Boltzmann Machines

– Deep Belief Networks– Deep Boltzmann Machines

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 16

Page 17: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning (Neural Networks) Approaches (Dates)

• Neural Network (Multi-Layer Perceptron) (1986)

• Convolutional Neural Networks (1989)

• Long-Short Term Memory LSTM (1997)

• Restricted Boltzmann Machines (2006)

– Deep Belief Networks– Deep Boltzmann Maniches

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 17

Page 18: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning (Neural Networks) approaches

• Supervised:

– DNN– Convolutional NN– LSTM

Training: Backpropagation

• Unsupervised:

– Stacked Autoencoders

Training: Backpropagation

– Restricted Boltzmann Machines

Contrastive Divergence or Persistent CD

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 18

Page 19: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning problems

• Problem when backpropagating errors to first layers

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 19

Page 20: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning problems

• Nowadays a DeepNet for computer vision is something like this:

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 20

Page 21: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning problems

• Backpropagation , Why now it works with deep structures?

– Tons of data– Sharing weights on very initial layers (CNN)– Improving Generalization:∗ Dropout∗ Dropconnect∗ Denoising Autoencoders

– New activation functions:∗ ReLU∗ MaxOut⇒ piecewise-linear behaviour and constant gradients

– Layer by layer training– Virtual Data with common distortions– Hardware allows to run experiments!! (GPU)

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 21

Page 22: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning things to consider

• Supervised training (DNN,CNN):

– Network topology– Sigmoid, tanh, ReLu, softmax, linear, Maxout– Data normalization– Virtual data– Batch size, Epochs– Weights initialization– Learning rate– Momentum rate– Dropout– Dropconnect– ...

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 22

Page 23: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning, when?

• When to use Deep Learning:

– There is an big gap between raw representation and categories– Hand-crafted features didn’t work– There are a lot of data for training (or virtual distortions)– Good hardware is available

• When hand-crafted features are good (expert knowledge):

– SVM– Random Forests, ERT– AdaBoost– ...

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 23

Page 24: Demystifying Deep Learning - Roberto Paredes Palacios @ PAPIs Connect

Deep Learning

Thanks for your attention

March 15, 2016 - PAPIs Connect, Valencia Deep Learning Page 24