How Microsoft Had Made Deep Learning Red-Hot in IT Industry Zhijie Yan, Microsoft Research Asia USTC visit, May 6, 2014 Self Introduction @MSRA鄢志杰 996 – studied in USTC from 1999 to 2008 Graduate student – studied in iFlytek speech lab from 2003 to 2008, supervised by Prof. Renhua Wang Intern – worked in MSR Asia from 2005 to 2006 Visiting scholar – visited Georgia Tech in 2007 FTE – worked in MSR Asia since 2008 Research interests Speech, deep learning, large-scale machine learning In Today’s Talk Deep learning becomes very hot in the past few years How Microsoft had made deep learning hot in IT industry Deep learning basics Why Microsoft can turn all these ideas into reality Further reading materials How Hot is Deep Learning “This announcement comes on the heels of a $600,000 gift Google awarded Professor Hinton’s research group to support further work in the area of neural nets.” – U. of T. website How Hot is Deep Learning How Hot is Deep Learning How Hot is Deep Learning How Hot is Deep Learning Microsoft Had Made Deep Learning Hot in IT Industry Initial attempts made by University of Toronto had shown promising results using DL in speech recognition on TIMIT phone recognition task Prof. Hinton’s student visited MSR as an intern, good results were obtained on Microsoft Bing voice search task MSR Asia and Redmond collaborated and got amazing results on Switchboard task, which shocked the whole industry Microsoft Had Made Deep Learning Hot in IT Industry *figure borrowed from MSR principal researcher Li DENG Microsoft Had Made Deep Learning Hot in IT Industry Followed by others and results were confirmed in various different speech recognition tasks Google / IBM / Apple / Nuance / 百度 / 讯飞 Continuously advanced by MSR and others Expand to solve more and more problems Image processing Natural language processing Search … Deep Learning From Speech to Image ILSVRC-2012 competition on ImageNet Classification task: classify an image into 1 of the 1,000 classes in your 5 bets lifeboat airliner school bus Institution Error rate (%) University of Amsterdam 29.6 XRCE/INRIA 27.1 Oxford 27.0 ISI 26.2 Deep Learning From Speech to Image ILSVRC-2012 competition on ImageNet Classification task: classify an image into 1 of the 1,000 classes in your 5 bets lifeboat airliner school bus Institution Error rate (%) University of Amsterdam 29.6 XRCE/INRIA 27.1 Oxford 27.0 ISI 26.2 SuperVision 16.4 Deep Learning Basics Deep learning deep neural networks multi-layer perceptron (MLP) with a deep structure (many hidden layers) Output layer Output layer W1 Hidden layer W0 Input layer W3 Hidden layer W2 Hidden layer W1 Hidden layer W0 Input layer Deep Learning Basics Sounds not new at all? Sounds familiar like you’ve learned in class? Things not change over the years Network topology / activation functions / … Backpropagation (BP) Things changed recently Data Big data General-purpose computing on graphics processing units (GPGPU) “A bag of tricks” accumulated over the years E.g. Deep Neural Network for Speech Recognition Three key components that make DNN-HMM work Many layers of nonlinear feature transformation Tied triphones as the basis units for HMM states Long window of frames *figure borrowed from MSR senior researcher Dong YU E.g. Deep Neural Network for Image Classification The ILSVRC-2012 winning solution *figure copied from Krizhevsky, et al., “ImageNet Classification with Deep Convolutional Neural Networks” Scale Out Deep Leaning Training speed was a major problem of DL Speech recognition model trained with 1,800-hour data (~650,000,000 vector frames) costs 2 weeks using 1 GPU Image classification model trained with ~1,000,000 figures costs 1 weeks using 2 GPUs* How to scale out if 10x, 100x training data becomes available? *Krizhevsky, et al., “ImageNet Classification with Deep Convolutional Neural Networks” DNN-GMM-HMM Joint work with USTC-MSRA Ph.D. program student, Jian XU (许健, 0510) The “DNN-GMM-HMM” approach for speech recognition* DNN as hierarchical nonlinear feature extractor, trained using a sub-set of training data GMM-HMM as acoustic model, trained using full data *Z.-J. Yan, Q. Huo, and J. Xu, “A scalable approach to using DNNderived features in GMM-HMM based acoustic modeling for LVCSR” DNN-GMM-HMM GMM-HMM modeling of DNN-derived features: combine the best of both worlds DNNderived features PCA HLDA Tied-state WE-RDLT MMI sequence training CMLLR unsupervised adaptation Experimental Results 300hr DNN (18k states, 7 hidden layers) + 2,000hr GMM-HMM (18k states)* Training time reduced from 2 weeks to 3-5 days Word Error Rate (%) DNN-HMM (CE) DNN-GMM-HMM (RDLT) DNN-GMM-HMM (MMI) DNN-GMM-HMM (UA) 15.4 14 14.7 10% WERR 13.8 12 15% WERR 13.1 10 *Z.-J. Yan, Q. Huo, and J. Xu, “A scalable approach to using DNNderived features in GMM-HMM based acoustic modeling for LVCSR” A New Optimization Method Joint work with USTC-MSRA Ph.D. program student, Kai Chen (陈凯, 0700) Using 20 GPUs, time needed to train a 1,800-hour acoustic model is cut from 2 weeks to 12 hours, without accuracy loss The magic is to be published We believe the scalability issue in DNN training for speech recognition is now solved! Why Microsoft Can Do All These Good Things Research Bridge the gap between academia and industry via our intern and visiting scholar programs Scale out from toy problems to real-world industry-scale applications Product team Solve practical issues and deploy technologies to serve users worldwide via our services All together We continuously improve our work towards larger scale, higher accuracy, and to tackle more challenging tasks Finally We have big-data + world-leading computational infrastructure If You Want to Know More About Deep Learning Neural networks for machine learning: https://class.coursera.org/neuralnets-2012-001 Prof. Hinton’s homepage: http://www.cs.toronto.edu/~hinton/ DeepLearning.net: http://deeplearning.net/ Open-source Kaldi (speech): http://kaldi.sourceforge.net/ cuda-convent (image): http://code.google.com/p/cuda-convnet/ Thanks!