![Classifier Preprocessing Test Error Rate (%) Reference
SVMs
SVM, Gaussian Kernel none 1.4
… … … …
Virtual SVM, deg-9 poly, 2-pixel jittered deskewing 0.56 DeCoste and Scholkopf, MLJ 2002
Neural Nets
2-layer NN, 300 hidden units, mean
square error
none 4.7 LeCun et al. 1998
…
6-layer NN 784-2500-2000-1500-1000-
500-10 (on GPU) [elastic distortions]
none 0.35
Ciresan et al. Neural Computation 10, 2010 and arXiv 1003.0358,
201
Convolutional nets
Convolutional net LeNet-1 subsampling to 16x16 pixels 1.7 LeCun et al. 1998
…
committee of 35 conv. net, 1-20-P-40-P-
150-10 [elastic distortions]
width normalization 0.23 Ciresan et al. CVPR 2012
Source: http://yann.lecun.com/exdb/mnist/
March 23, 2017 Djamal Abide 29](https://image.slidesharecdn.com/deepleaningcomputervision-170818034619/85/Application-of-deep-leaning-to-computer-vision-29-320.jpg)
Application of deep leaning to computer vision
- 1. Application of deep learning to computer vision Presented by: Djamal Abide
- 2. Plan 1. Data science 2. Artificial intelligence 3. Computer vision 4. Deep Learning 5. Demo March 23, 2017 Djamal Abide 2
- 3. Plan 1. Data science 2. Artificial intelligence 3. Computer vision 4. Deep Learning 5. Demo March 23, 2017 Djamal Abide 3
- 4. Data Science Definition It’s is an interdisciplinary field about processes and systems to extract knowledge or insights from data in various forms, either structured or unstructured March 23, 2017 Djamal Abide 4
- 5. Examples of AI Applications Type Examples Monitoring 1. Detecting credit-card fraud 2. Cybersecurity intrusions Discovering 1. Genetics 2. Causal models for air transport safety Predicting 1. Netflix movies recommendation 2. Weather forecasting Interpreting 1. Face detection (images) 2. Pedestrian detection (videos) 3. Speech recognition (audios) March 23, 2017 Djamal Abide 5
- 6. Data Science Data Engineering Scientific Method Math StatisticsAdvanced Computing Visualization Hacker Mindset Data Science Team Skills Set March 23, 2017 Djamal Abide 6
- 7. Ask Questions Research & Gather Data Formulate Hypothesis Test Hypothesis (Experiments) Analyze Results (Draw Conclusion) Report Results The Scientific Method March 23, 2017 Djamal Abide 7
- 8. Plan 1. Data science 2. Artificial intelligence 3. Computer vision 4. Deep Learning 5. Demo March 23, 2017 Djamal Abide 8
- 9. Artificial Intelligence Natural Language Processing (NLP) Computer Vision Robotics Problem-solving and planning Machine Learning Knowledge Representation Artificial Intelligence Research Fields March 23, 2017 Djamal Abide 9
- 10. Plan 1. Data science 2. Artificial intelligence 3. Computer vision 4. Deep Learning 5. Demo March 23, 2017 Djamal Abide 10
- 11. What is Computer Vision? It’s a field that includes methods for acquiring, processing, analyzing and understanding images from the real world in order to produce information in the form of decision. Applications • Recognize objects • Locate objects in space • Track objects • Recognize actions March 23, 2017 Djamal Abide 11
- 12. Computer Vision Optics Machine Learning Digital Images Processing Computer Vision Components March 23, 2017 Djamal Abide 12
- 13. Source: https://www.pro-therm.com/images/infrared_basics_figure2_large.gif Radiation wavelengths March 23, 2017 Djamal Abide 13
- 14. Colored Image Data Structure Red, Green and Blue values are between: 0 and 255 Intensity values are between: 0 and 255 Gray Scaled Image Data Structure March 23, 2017 Djamal Abide 14
- 15. Image Processing Examples Resized Gray Scale Edge Detection March 23, 2017 Djamal Abide 15
- 16. Classical Program x f(x) y Machine Learning: f(x) function is Learned from the data Machine Learning vs Classical Program Input Data Program Implementing f(x) Result (x1, y1) (x2, y2) ... ML Algorithm Model f(x) Training Examples Result ML Program To Learn f(x) March 23, 2017 Djamal Abide 16
- 17. Prediction Evaluation Prediction With Machine Learning Model Model f(x) x Prediction Tool Predicted y Predicted y Real y Comparison Tool Accuracy March 23, 2017 Djamal Abide 17
- 18. March 23, 2017 Djamal Abide 18
- 19. Plan 1. Data science 2. Artificial intelligence 3. Computer vision 4. Deep Learning 5. Demo March 23, 2017 Djamal Abide 19
- 20. Source: https://i.ytimg.com/vi/osa3zIEJjgw/maxresdefault.jpg Human brain and Artificial Neural Networks Human brain doesn’t need features Activation function March 23, 2017 Djamal Abide 20
- 21. Source: https://nivdul.files.wordpress.com/2015/11/nivdul_deep_learning.png?w=700&h=367 Deep neural networks learn hierarchical feature representations March 23, 2017 Djamal Abide 21
- 22. Deep Learning Flow For Training Models Input data Preprocessing Enhanced Clean Data Features Extraction Features Deep Learning Model • Without clean data, Deep Learning cannot learn or discover patterns Traditional Machine Learning Flow For Training Models Input data Preprocessing Enhanced Clean Data Features Extraction Features (help in finding patterns) Tradition ML Algorithm Model • Clean data helps in engineering robust features • Without good features, ML algorithm cannot learn or discover patterns X X March 23, 2017 Djamal Abide 22
- 23. Why it is hard to recognize objects? • Segmentation: Picture contains many objects • Lighting: Intensity of light • Deformation: Handwriting with many styles • Affordance: Objects labeled based on what they are used for. Example: chairs • Viewpoint: Picture could be taken from different angles March 23, 2017 Djamal Abide 23
- 24. Convolutional layer March 23, 2017 Djamal Abide 24
- 25. Pooling layer March 23, 2017 Djamal Abide 25
- 26. LeNet: 1st successful CNN March 23, 2017 Djamal Abide 26 Source: http://yann.lecun.com/exdb/publis/pdf/lecun-98.pdf
- 27. • MNIST has contains 70,000 of pictures 9 different digits • Format of a picture is 28 x 28 • Scientists use 60,000 pics to train and 10,000 pics for testing MNIST Database March 23, 2017 Djamal Abide 27
- 28. Classifier Preprocessing Test Error Rate (%) Reference Linear Classifiers linear classifier (1-layer NN) none 12.0 LeCun et al. 1998 linear classifier (1-layer NN) deskewing 8.4 LeCun et al. 1998 pairwise linear classifier deskewing 7.6 LeCun et al. 1998 K-Nearest Neighbors K-nearest-neighbors, Euclidean (L2) none 5.0 LeCun et al. 1998 … K-NN, shape context matching shape context feature extraction 0.63 Belongie et al. IEEE PAMI 2002 Boosted Stumps boosted stumps none 7.7 Kegl et al., ICML 2009 … product of stumps on Haar features Haar features 0.87 Kegl et al., ICML 2009 Non-Linear Classifiers 40 PCA + quadratic classifier none 3.3 LeCun et al. 1998 1000 RBF + linear classifier none 3.6 LeCun et al. 1998 March 23, 2017 Djamal Abide 28
- 29. Classifier Preprocessing Test Error Rate (%) Reference SVMs SVM, Gaussian Kernel none 1.4 … … … … Virtual SVM, deg-9 poly, 2-pixel jittered deskewing 0.56 DeCoste and Scholkopf, MLJ 2002 Neural Nets 2-layer NN, 300 hidden units, mean square error none 4.7 LeCun et al. 1998 … 6-layer NN 784-2500-2000-1500-1000- 500-10 (on GPU) [elastic distortions] none 0.35 Ciresan et al. Neural Computation 10, 2010 and arXiv 1003.0358, 201 Convolutional nets Convolutional net LeNet-1 subsampling to 16x16 pixels 1.7 LeCun et al. 1998 … committee of 35 conv. net, 1-20-P-40-P- 150-10 [elastic distortions] width normalization 0.23 Ciresan et al. CVPR 2012 Source: http://yann.lecun.com/exdb/mnist/ March 23, 2017 Djamal Abide 29
- 30. Deep Learning: GPU versus CPU March 23, 2017 Djamal Abide 30 Source: http://www.nvidia.com/object/tesla-m40.html
- 31. Large Scale Visual Recognition Challenge 2012 (ILSVRC2012) • Number of images: ~ 14 million • Number of categories: 1,000 • Team “SuperVision” formed by students of Professor Geoffrey Hinton from University of Toronto Alex Krizhevsky & Ilya Sutskever won ImageNet classification challenge with a large margin March 23, 2017 Djamal Abide 31
- 32. Pros • Enable learning of features rather than hand tuning • Impressive performance gains in: – Computer vision – Speech recognition – Some text analysis • Potential for more impact Cons • Requires a lot of data for high accuracy • Computationally really expensive • Hard to tune: – Choice of architecture – Parameter types – Hyper-parameters – Learning algorithm – … Deep Learning: Pros & Cons March 23, 2017 Djamal Abide 32
- 33. Advise • Use segmented images as training set • Use data augmentation technics • Don’t be a ‘hero’ trying to create your own Deep Neuronal Network (CNN) architecture, use an existing one • Use transfer learning (pre-trained models) March 23, 2017 Djamal Abide 33
- 34. Plan 1. Data science 2. Artificial intelligence 3. Computer vision 4. Deep Learning 5. Demo March 23, 2017 Djamal Abide 34
- 35. ConvNetJS (Deep Learning in your browser) • http://cs.stanford.edu/people/karpathy/convn etjs/index.html March 23, 2017 Djamal Abide 35
- 36. March 23, 2017 Djamal Abide 36