EuroScipy2014

Abstract here: https://www.euroscipy.org/2014/schedule/presentation/2/

Code: https://github.com/sklearn-theano/sklearn-theano

Transcript

1. Neural Networks For Computer Vision Kyle Kastner INRIA Parietal Université de Montréal
2. Neural Networks For Computer Vision Kyle Kastner INRIA Parietal Université de Montréal
3. The Most Dangerous Game: Sloth Hunting 101 Kyle Kastner INRIA Parietal Université de Montréal
4. The Most Dangerous Game: Sloth Hunting 101 Kyle Kastner INRIA Parietal Université de Montréal * No animals were harmed in making this talk ** Please don’t hunt sloths
5. The Plan Classify Locate ??? Profit
6. Computer Vision Fixed Feature Extractors Simple Classifier SURF SIFT SVM HOG skimage sklearn HOG linear models
7. “Deep” Computer Vision Trainable Feature Extractor Trainable Classifier Trained together theano pylearn2 Not pictured: 1.2 TB of images (ImageNet, 1.2E6 images) compute cluster / monster GPU(s) neural network neural network
8. Concepts • Neural networks are universal function approximators • Can train these to do (almost) whatever we want ◦ Will it generalize to new data? • Computer Vision: image descriptors • Machine Learning: features
9. Exploiting Structure [r8] conv pool conv pool [1],[2],[3]
10. Technology • Theano • Optimizing compiler • Lots of mathematical tools • Many operations from numpy “just work” • Automatic differentiation • GPU or CPU set by single flag ◦ THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32
11. Cherry On Top • pylearn2 • Built with Theano • Used for neural network research • Great guides to YAML interface • Can code in pure python Monitoring step: Epochs seen: 33 Batches seen: 16500 Examples seen: 1650000 learning_rate: 0.00999999046326 momentum: 0.989998817444 test_h2_kernel_norms_max: 1.87118041515 test_h2_kernel_norms_mean: 0.804925084114 test_h2_kernel_norms_min: 0.109122686088 test_h3_kernel_norms_max: 1.90657293797 test_h3_kernel_norms_mean: 1.3938267231 test_h3_kernel_norms_min: 0.348352521658 test_objective: 0.0377874299884 test_term_0: 0.0261260885745 test_term_1_weight_decay: 0.0116613674909 test_y_max_max_class: 0.999999344349 test_y_mean_max_class: 0.99490904808 test_y_min_max_class: 0.76411986351 test_y_misclass: 0.00859999842942 valid_h2_kernel_norms_max: 1.87118041515 valid_h2_kernel_norms_mean: 0.804925084114 valid_h2_kernel_norms_min: 0.109122686088 valid_h3_kernel_norms_max: 1.90657293797 valid_h3_kernel_norms_mean: 1.3938267231 valid_h3_kernel_norms_min: 0.348352521658 valid_objective: 0.0444900207222 valid_term_0: 0.0328286737204 valid_term_1_weight_decay: 0.0116613674909 valid_y_max_max_class: 0.999999344349 valid_y_mean_max_class: 0.995154678822 valid_y_min_max_class: 0.737455904484 valid_y_misclass: 0.00819999724627 https://github.com/kastnerkyle/pylearn2-practice
12. sklearn-theano • Michael Eickenberg and myself • Easy interface to trained networks • Other Theano based models • (Mostly) scikit-learn compatible • Includes examples from this talk https://github.com/sklearn-theano/sklearn-theano
13. sklearn-theano Trained Feature Extractor Trainable Classifier Canny sklearn pylearn2 models OverFeat (NYU)[4] DeCAF (UCB) Canny Not pictured: Weeks of expert time and compute time Cross-validation of entire networks! linear models SVM
14. Classification sklearn_theano/examples/plot_classification.py
15. Localization sklearn_theano/examples/plot_localization_tutorial.py
16. Generative Adversarial Networks • Ian Goodfellow and a slew of others, 2014 [5] • Trains two networks ◦ Generate examples from random input ◦ Discriminate whether generated or real Generation sklearn_theano/examples/plot_mnist_generator.py
17. Takeaways • Neural networks are good for many vision tasks! • Theano is a great package for {C,G}PU math in Python • pylearn2 is an awesome neural net package! • Use large neural nets easily in sklearn-theano! • Not every problem needs a massive neural network! • But some do!
18. Slides: https://speakerdeck.com/kastnerkyle/euroscipy2014 https://github.com/kastnerkyle/EuroScipy2014 Code: https://github.com/sklearn-theano/sklearn-theano Thank You!
19. Learning Materials: Courses Coursera ML course, Andrew Ng https://www.coursera.org/course/ml This course is phenomenal, and will force you to learn machine learning including simple neural nets. Yes, it is MATLAB, sorry. Still worth it. Coursera Neural Networks Course, Geoff Hinton https://www.coursera.org/course/neuralnets This course is a detailed look at modern neural networks by one of the core researchers in the field. Well worth it, and goes up to the edge of research. Deep Learning Tutorial, LISA Lab, Yoshua Bengio http://deeplearning.net/tutorial/ If you want to understand Theano and how to build neural nets, this is an excellent tutorial. After this you will understand the whole of my talk! Neural Networks Course, Hugo Larochelle https://www.youtube.com/playlist?list=PL6Xpj9I5qXYEcOhn7TqghAJ6NAPrNmUBH This course will lock down every single piece of neural network approaches, and covers many important topics.
20. Learning Materials: Papers Representation Learning, A Review and New Perspectives, Yoshua Bengio, Aaron Courville, Pascal Vincent http://arxiv.org/abs/1206.5538 This paper covers much of the “nitty gritty” details. Weight initialization, optimization schemes, tradeoffs, and implementation. Fantastic! Learning Features from Tiny Images, Alex Krizhevsky http://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf The greedy layerwise training of this thesis is out of date now, but the rest of the insights into image statistics and intuition are very, very good. ImageNet Classification with Deep Convolutional Neural Networks, Alex Krizhevsky, Ilya Sutskever, Geoff Hinton http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks This paper basically lays out the “modern” approach to most neural networks for images (also audio, and some NLP). OverFeat is basically an implementation of this architecture, with a number of custom tweaks to get better performance. OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks, P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, Y. LeCun http://arxiv.org/abs/1312.6229 I like this paper (and associated code: https://github.com/sermanet/OverFeat) a whole lot. It inspired sklearn-theano!
21. Learning Materials: Blogs Kaggle Galaxy Challenge Solution, Sander Dieleman http://benanne.github.io/2014/04/05/galaxy-zoo.html This blog post will show you how to build a neural network architecture from scratch for a custom dataset. His post on audio for Spotify is also spectacular: http://benanne.github.io/2014/08/05/spotify-cnns.html Deep Learning in One File, Gabriel Synnaeve http://snippyhollow.github.io/ This blog post has great, straightforward insight into deep learning, and provides a single file solution to do deep learning using Theano! Neural Networks, Manifolds, and Topology, Chris Olah http://colah.github.io/posts/2014-03-NN-Manifolds-Topology/ A whole blog full of remarkable discussions of neural networks, but this post gives some neat intuition about how and why neural networks work. This post on NLP is also great: http://colah.github.io/posts/2014-07-NLP-RNNs-Representations/ Using Deep Learning To Listen For Whales, Daniel Nouri http://danielnouri.org/notes/2014/01/10/using-deep-learning-to-listen-for-whales/ This model did not win, but the level of effort the winner put into feature engineering was immense. This approach did very well, very simply. Awesome!
22. References [1] http://deeplearning.stanford.edu/wiki/index.php/UFLDL_Tutorial [2] http://www.cs.nyu.edu/~yann/talks/lecun-ranzato-icml2013.pdf [3] http://parse.ele.tue.nl/education/cluster2 [4] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, Y. LeCun. OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks, International Conference on Learning Representations (ICLR 2014), April 2014. http://arxiv.org/abs/1406.2661 [5] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio. Generative Adversarial Networks, arXiV, June 2014. http://arxiv.org/abs/1406.2661
23. BONUS SLIDES
24. Preprocessing class precision recall f1-score support 0.0 0.89 0.89 0.89 9 1.0 0.91 0.91 0.91 11 avg / total 0.90 0.90 0.90 20 accuracy score 0.9 tf = OverfeatTransformer(output_layers=[-3]) clf = LogisticRegression() pipe = make_pipeline(tf, clf) pipe.fit(X_train, y_train) y_pred = pipe.predict(X_test) Code sklearn-theano/examples/asirra_dataset_example.py
25. Data Werewolves... • But no silver bullets! • Primarily for automation (WHAT) • Other models can be better for ◦ Scientific inference (WHY) ◦ Simplicity of tuning (HOW) ◦ Underpowered hardware (WHO)