Empfohlen
Weitere ähnliche Inhalte
Was ist angesagt?
Was ist angesagt? (14)
Mehr von zukun
Mehr von zukun (20)
Kürzlich hochgeladen
Kürzlich hochgeladen (20)
Quoc le tera-scale deep learning
- 1. Tera-scale deep learning Quoc V. Le Stanford University and Google
- 2. Joint work with Kai Chen Greg Corrado Jeff Dean MaQhieu Devin Rajat Monga Andrew Ng Marc Aurelio Paul Tucker Ke Yang Ranzato Samy Bengio, Zhenghao Chen, Tom Dean, Pangwei Koh, AddiNonal Mark Mao, Jiquan Ngiam, Patrick Nguyen, Andrew Saxe, Thanks: Mark Segal, Jon Shlens, Vincent Vanhouke, Xiaoyun Wu, Peng Xe, Serena Yeung, Will Zou
- 3. Machine Learning successes Face recogniNon OCR Autonomous car Email classificaNon RecommendaNon systems Web page ranking
- 4. Feature ExtracNon Classifier Feature extracNon (Mostly hand-‐cra]ed features)
- 5. Hand-‐Cra]ed Features Computer vision: … SIFT/HOG SURF Speech RecogniNon: … MFCC Spectrogram ZCR
- 6. New feature-‐designing paradigm Unsupervised Feature Learning / Deep Learning ReconstrucNon ICA Expensive and typically applied to small problems
- 7. The Trend of BigData
- 8. Outline No maQer the algorithm, more features always more successful. -‐ ReconstrucNon ICA -‐ ApplicaNons to videos, cancer images -‐ Ideas for scaling up -‐ Scaling up Results
- 9. Topographic Independent Component Analysis (TICA)
- 10. 1. Feature computaNon 2 2. Learning 2 ( W T ( W T ) 9 ) 1 W T W T 9 1 W W 9 1 W 1 W = W 2 . . W 10000 Input data:
- 11. Topographic Independent Component Analysis (TICA)
- 12. Invariance explained Images Image1 Image2 Features Loc1 Loc2 1 0 F1 F2 0 1 Pooled feature of F1 and F2 sqrt(1 + 02 ) = 1 2 sqrt(0 2 + 12 ) = 1 Same value regardless the locaNon of the edge
- 13. TICA: ReconstrucNon ICA: Equivalence between Sparse Coding, Autoencoders, RBMs and ICA Build deep architecture by treaNng the output of one layer as input to another layer Le, et al., ICA with Reconstruc1on Cost for Efficient Overcomplete Feature Learning. NIPS 2011
- 14. ReconstrucNon ICA: Le, et al., ICA with Reconstruc1on Cost for Efficient Overcomplete Feature Learning. NIPS 2011
- 15. ReconstrucNon ICA: Data whitening Le, et al., ICA with Reconstruc1on Cost for Efficient Overcomplete Feature Learning. NIPS 2011
- 16. TICA: ReconstrucNon ICA: Data whitening Le, et al., ICA with Reconstruc1on Cost for Efficient Overcomplete Feature Learning. NIPS 2011
- 17. Why RICA? Algorithms Speed Ease of training Invariant Features Sparse Coding RBMs/Autoencoders TICA ReconstrucNon ICA Le, et al., ICA with Reconstruc1on Cost for Efficient Overcomplete Feature Learning. NIPS 2011
- 18. Summary of RICA -‐ Two-‐layered network -‐ ReconstrucNon cost instead of orthogonality constraints -‐ Learns invariant features
- 20. AcNon recogniNon Sit up Drive Car Get Out of Car Eat Answer phone Kiss Run Stand up Shake hands Le, et al., Learning hierarchical spa1o-‐temporal features for ac1on recogni1on with independent subspace analysis. CVPR 2011
- 21. Le, et al., Learning hierarchical spa1o-‐temporal features for ac1on recogni1on with independent subspace analysis. CVPR 2011
- 22. 94 55 KTH 53 Hollywood2 92 51 90 49 88 47 45 86 43 84 41 39 82 37 80 35 Hessian/SURF pLSA HOF GRBMs 3DCNN HMAX HOG Hessian/SURF HOG/HOF HOG3D GRBMS HOF Learned Features Learned Features 87 76 UCF YouTube 85 75 83 74 81 73 79 72 77 71 75 70 Hessian/SURF HOG Hessian HOF HOG3D Combined Learned Features HOG.HOF Learned Features Engineered Features Le, et al., Learning hierarchical spa1o-‐temporal features for ac1on recogni1on with independent subspace analysis. CVPR 2011
- 23. Cancer classificaNon 92% ApoptoNc 90% Viable tumor 88% region 86% Necrosis 84% Hand engineered Features RICA … Le, et al., Learning Invariant Features of Tumor Signatures. ISBI 2012
- 24. Scaling up deep RICA networks
- 25. Scaling up Deep Learning Deep learning data Real data
- 26. It’s beQer to have more features! No maQer the algorithm, more features always more successful. Coates, et al., An Analysis of Single-‐Layer Networks in Unsupervised Feature Learning. AISTATS’11
- 27. Most are local features
- 28. Local recepNve field networks Machine #1 Machine #2 Machine #3 Machine #4 RICA features Image Le, et al., Tiled Convolu1onal Neural Networks. NIPS 2010
- 29. Challenges with 1000s of machines
- 30. Asynchronous Parallel SGDs Parameter server Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 31. Asynchronous Parallel SGDs Parameter server Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 32. Summary of Scaling up -‐ Local connecNvity -‐ Asynchronous SGDs … And more -‐ RPC vs MapReduce -‐ Prefetching -‐ Single vs Double -‐ Removing slow machines -‐ OpNmized So]max -‐ …
- 33. 10 million 200x200 images 1 billion parameters
- 34. Training RICA Dataset: 10 million 200x200 unlabeled images from YouTube/Web Train on 2000 machines (16000 cores) for 1 week RICA 1.15 billion parameters -‐ 100x larger than previously reported -‐ Small compared to visual cortex RICA Image Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 35. The face neuron Top sNmuli from the test set OpNmal sNmulus by numerical opNmizaNon Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 36. Random distractors Faces Frequency Feature value Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 37. Feature response Invariance properNes Feature response 0 pixels 20 pixels 0 pixels 20 pixels Horizontal shi]s VerNcal shi]s Feature response Feature response o o 0 90 0.4x 1x 1.6x 3D rotaNon angle Scale factor Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 38. Top sNmuli from the test set OpNmal sNmulus by numerical opNmizaNon Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 39. Random distractors Pedestrians Frequency Feature value Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 40. Top sNmuli from the test set OpNmal sNmulus by numerical opNmizaNon Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 41. Random distractors Cat faces Frequency Feature value Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 43. ImageNet classificaNon 22,000 categories 14,000,000 images Hand-‐engineered features (SIFT, HOG, LBP), SpaNal pyramid, SparseCoding/Compression Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 44. 22,000 is a lot of categories… … smoothhound, smoothhound shark, Mustelus mustelus American smooth dogfish, Mustelus canis Florida smoothhound, Mustelus norrisi whiteNp shark, reef whiteNp shark, Triaenodon obseus AtlanNc spiny dogfish, Squalus acanthias Pacific spiny dogfish, Squalus suckleyi SNngray hammerhead, hammerhead shark smooth hammerhead, Sphyrna zygaena smalleye hammerhead, Sphyrna tudes shovelhead, bonnethead, bonnet shark, Sphyrna Nburo angel shark, angelfish, SquaNna squaNna, monkfish electric ray, crampfish, numbfish, torpedo Mantaray smalltooth sawfish, PrisNs pecNnatus guitarfish roughtail sNngray, DasyaNs centroura buQerfly ray eagle ray spoQed eagle ray, spoQed ray, Aetobatus narinari cownose ray, cow-‐nosed ray, Rhinoptera bonasus manta, manta ray, devilfish AtlanNc manta, Manta birostris devil ray, Mobula hypostoma grey skate, gray skate, Raja baNs liQle skate, Raja erinacea …
- 45. Best sNmuli Feature 1 Feature 2 Feature 3 Feature 4 Feature 5 Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 46. Best sNmuli Feature 6 Feature 7 Feature 8 Feature 9 Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 47. Best sNmuli Feature 10 Feature 11 Feature 12 Feature 13 Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 48. 0.005% 9.5% ? Random guess State-‐of-‐the-‐art Feature learning (Weston, Bengio ‘11) From raw pixels Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 49. 0.005% 9.5% 15.8% Random guess State-‐of-‐the-‐art Feature learning (Weston, Bengio ‘11) From raw pixels ImageNet 2009 (10k categories): Best published result: 17% (Sanchez & Perronnin ‘11 ), Our method: 20% Using only 1000 categories, our method > 50% Le, et al., Building high-‐level features using large-‐scale unsupervised learning. ICML 2012
- 50. Other results No maQer the algorithm, more features always more successful. -‐ We also have great features for -‐ Speech recogniNon -‐ Word-‐vector embedding for NLPs
- 51. Conclusions • RICA learns invariant features • Face neuron with totally unlabeled data ImageNet with enough training and data • State-‐of-‐the-‐art performances on 0.005% 9.5% 15.8% Random guess Best published result Our method – AcNon RecogniNon – Cancer image classificaNon – ImageNet 94 92 90 88 86 84 82 80 Cancer classificaNon AcNon recogniNon AcNon recogniNon benchmarks Feature visualizaNon Face neuron
- 52. References • Q.V. Le, M.A. Ranzato, R. Monga, M. Devin, G. Corrado, K. Chen, J. Dean, A.Y. Ng. Building high-‐level features using large-‐scale unsupervised learning. ICML, 2012. • Q.V. Le, J. Ngiam, Z. Chen, D. Chia, P. Koh, A.Y. Ng. Tiled Convolu8onal Neural Networks. NIPS, 2010. • Q.V. Le, W.Y. Zou, S.Y. Yeung, A.Y. Ng. Learning hierarchical spa8o-‐temporal features for ac8on recogni8on with independent subspace analysis. CVPR, 2011. • Q.V. Le, J. Ngiam, A. Coates, A. Lahiri, B. Prochnow, A.Y. Ng. On op8miza8on methods for deep learning. ICML, 2011. • Q.V. Le, A. Karpenko, J. Ngiam, A.Y. Ng. ICA with Reconstruc8on Cost for Efficient Overcomplete Feature Learning. NIPS, 2011. • Q.V. Le, J. Han, J. Gray, P. Spellman, A. Borowsky, B. Parvin. Learning Invariant Features for Tumor Signatures. ISBI, 2012. • I.J. Goodfellow, Q.V. Le, A.M. Saxe, H. Lee, A.Y. Ng, Measuring invariances in deep networks. NIPS, 2009. hQp://ai.stanford.edu/~quocle