Open source ai_technical_trend
2 likes753 views
This document discusses artificial intelligence and machine learning technologies. It provides an overview of machine learning and neural networks. It then discusses various open-source AI software frameworks including Caffe, Chainer, CNTK, DL4J, Keras, MXNet, Theano, Torch, and TensorFlow. For each framework, it provides information on the creator, applications, and motivation. It also includes the developer's experience working with image recognition, medical data processing, and various open-source projects.
![• Created & Maintained by
Preferred Networks, Inc.
(https://www.preferred-networks.jp/ja/)
• Release
Jun. „2015
• Application
Toyota motors, Panasonic
(https://www.wsj.com/articles/japan-seeks-tech-revival-with-artificial-intelligence-
1448911981)
FANUC
(http://www.fanucamerica.com/FanucAmerica-news/Press-
releases/PressReleaseDetails.aspx?id=79)
• Motivation
Define-by-Run Architecture
Chainer
http://docs.chainer.org/en/latest/index.html
S/W Creator Platform Mobile
Langua
ge
Interface OpenMP CUDA OpenCL Multi GPU
Parallel
Executi
on
Chainer
Preferred
Networks
Linux - Python Python -
Y
- Y Y
[Define-and-Run (TensorFlow)] [Define-by-Run (Chainer, PyTorch)]](https://image.slidesharecdn.com/opensourceaitechnicaltrend-171121051204/85/Open-source-ai_technical_trend-30-320.jpg)
![Edges2Image
[Isola et al. CVPR 2017]](https://image.slidesharecdn.com/opensourceaitechnicaltrend-171121051204/85/Open-source-ai_technical_trend-69-320.jpg)
![3D Generative Adversarial Network
[Wu et al. NIPS 2016]](https://image.slidesharecdn.com/opensourceaitechnicaltrend-171121051204/85/Open-source-ai_technical_trend-71-320.jpg)
![Scene Parsing
[Farabet et al. ICML 2012, PAMI 2013]](https://image.slidesharecdn.com/opensourceaitechnicaltrend-171121051204/85/Open-source-ai_technical_trend-73-320.jpg)
![Scene Parsing
[Farabet et al. ICML 2012, PAMI 2013]](https://image.slidesharecdn.com/opensourceaitechnicaltrend-171121051204/85/Open-source-ai_technical_trend-74-320.jpg)
Open source ai_technical_trend
- 1. 인공지능 (A.I.) 이해와 최신기술 Artificial Intelligence and technical trend. Lablup Inc. Mario Cho (조만석) hephaex@gmail.com
- 2. Contents • Machine Learning? • Artificial Neural Network? • Open Source based Artificial Intelligence Softwares • Open Source A.I Software Applications
- 3. Mario (manseok) Cho Development Experience Image Recognition using Neural Network Bio-Medical Data Processing Human Brain Mapping on High Performance Computing Medical Image Reconstruction(Computer Tomography) Enterprise System Architect & consuliting Artificial Intelligence for medicine decision support Open Source Software Developer Committer: (Cloud NFV/SDN) Contribute: TensorFlow (Deep Learning) OpenStack (Cloud compute) LLVM (compiler) Kernel (Linux) Book Unix V6 Kernel Lablup Inc. Mario Cho hephaex@gmail.com
- 4. The Future of Jobs “The Fourth Industrial Revolution, which includes developments in previously disjointed fields such as artificial intelligence & machine-learning, robotics, nanotechnology, 3-D printing, and genetics & biotechnology, will cause widespread disruption not only to business models but also to labor market over the next five years, with enormous change predicted in the skill sets needed to thrive in the new landscape.”
- 6. What is the Machine Learning ? • Field of Computer Science that evolved from the study of pattern recognition and computational learning theory into Artificial Intelligence. • Its goal is to give computers the ability to learn without being explicitly programmed. • For this purpose, Machine Learning uses mathematical / statistical techniques to construct models from a set of observed data rather than have specific set of instructions entered by the user that define the model for that set of data.
- 7. Artificial Intelligence Understand information, To Learn, To Reason, & Act upon it
- 9. What is a neural network? Yes/No (Mug or not?) Data (image) x1 λ 5 , x2 λ 5 x2 =(W1 ´x1 )+ x3 =(W2 ´x2 )+ x1 x2 x3 x4 x5 W4W3W2W1
- 10. Neural network vs Learning network Neural Network Deep Learning Network
- 12. Neural Network 2.7 -8.6 0.002 f(x) 1.4 -2.5 -0.06 x = -0.06×2.7 + 2.5×8.6 + 1.4×0.002 = 21.34
- 13. Neural Network as a Computational Graph • In Most Machine Learning Frameworks, • Neural Network is conceptualized as a Computational Graph • The simple form of Computational Graph, • Directed Acyclic Graph consist Data Nodes and Operator Nodes Y = x1 * x2 Z = x3 – y Data node Opeator node
- 14. Computational Graph Tensor (Muti axis matrix) Tensor Tensor Multifly add Function
- 15. Single layer perceptron Affine ReLUX W b h1 C C = ReLU( b + WX )
- 16. Multi layer perceptron X W1 b1 h1Affine a1 W2 b2 h2Affine ReLU ReLU a2 W3 b3 h3Affine Softmax t Cross Entropy prob loss
- 17. WFO Discovery Advisor • Researches can‟t innovate fast enough to create truly breakthrough therapies • To anticipate the safety profile of new treatments WFO Corpus Over 1TB of data Over 40m documents Over 100m entities & relationships Chemical 12M+ Chemical Structures Genomics 20,000+ genes MD Text 50+ books Medline 23M+ abstracts Journals 100+ journals FDA drugs 11,000+ drugs Patents 16M+ patents
- 18. GPU
- 19. Tensor Core : NVIDIA Volta
- 20. Why is Deep Learning taking off? Engine Fuel Large neural networks Labeled data (x,y pairs)
- 22. Deep learning : CNN
- 23. Traditional learning vs Deep Machine Learning Eiffel Tower Eiffel Tower RAW data RAW data Deep Learning Network Feature Extraction Vectored Classification Traditional Learning Deep Learning
- 24. Human-Level Object Recognition • ImageNet • Large-Scale Visual Recognition Challenge Image Classification / Localization 1.2M labeled images, 1000 classes Convolutional Neural Networks (CNNs) has been dominating the contest since.. 2012 non-CNN: 26.2% (top-5 error) 2012: (Hinton, AlexNet)15.3% 2013: (Clarifai) 11.2% 2014: (Google, GoogLeNet) 6.7% 2015: (Google) 4.9% Beyond human-level performance
- 25. The Big Players
- 26. History of Deep Learning Framework 2010 2013 2014 2015 2016 2017 (Nov.) (Dec.) (Jul.) (Jun.) On GitHub (Debut: Apr. ‘2015) (Oct.) (Jun.) (Nov.) (Jan.) (Apr.) (Mar.)
- 27. Google
- 28. Open Source Software for Machine Learning Caffe Theano Convnet.js Torch7 Chainer DL4J TensorFlow Neon SANOA Summingbird Apache SA Flink ML Mahout Spark MLlib RapidMiner Weka Knife Scikit-learn Amazon ML BigML DataRobot FICO Google prediction API HPE haven OnDemand IBM Watson PurePredictive Yottamine Deep Learning Stream Analytics Big Data Machine Learning Data Mining Machine Learning As a Service Pylearn2
- 29. • Created by Yangqing Jia (http://daggerfs.com/) UC Berkerey Computer Science Ph.D. / Trevor Darrell, BAIR Google BrainLab.TensorFlow join Facebook research Scientest Evan Shellhamer (http://imaginarynumber.net/) • Maintained by BAIR(Berkeley Artificial Intelligence Research, http://bair.berkeley.edu/) • Release „2013: DeCAF (https://arxiv.org/abs/1310.1531) Dec. „2013: Caffe v0 • Application Facebook, Adobe, Microsoft, Samsung, Flickr, Tesla, Yelp, Pinterest, etc. • Motivation „2012 ILSVRC, AlexNet DNN define/training/deploy implementation by F/W Caffe http://caffe.berkeleyvision.org/ S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Executi on Caffe BAIR Linux, Mac - C++ Python, MATLAB Y Y - Y
- 30. • Created & Maintained by Preferred Networks, Inc. (https://www.preferred-networks.jp/ja/) • Release Jun. „2015 • Application Toyota motors, Panasonic (https://www.wsj.com/articles/japan-seeks-tech-revival-with-artificial-intelligence- 1448911981) FANUC (http://www.fanucamerica.com/FanucAmerica-news/Press- releases/PressReleaseDetails.aspx?id=79) • Motivation Define-by-Run Architecture Chainer http://docs.chainer.org/en/latest/index.html S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Executi on Chainer Preferred Networks Linux - Python Python - Y - Y Y [Define-and-Run (TensorFlow)] [Define-by-Run (Chainer, PyTorch)]
- 31. • Created & Maintained by Microsoft Research • Release Jan. „2016 • Applications Microsoft‟s speech recognition engine Skype‟s Translator • Motivation Efficient performance on distributed environments CNTK https://www.microsoft.com/en-us/research/product/cognitive-toolkit/ https://www.microsoft.com/en-us/research/blog/microsoft-computational-network-toolkit-offers-most-efficient-distributed-deep-learning-computational-performance/ S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Execution CNTK Microsoft Linux, Windows - C++ Python, C++ Y Y - Y Y
- 32. • 주체 • Created by Adam Gibson @Skymind (CTO) Chris Nicholson @Skymind (CEO) • Maintained by Skymind (https://skymind.ai/) • Release Jun. „2014 • Application Finatial Fraud Detection Research Partnership with Nextremer in Japan (https://skymind.ai/press/nextremer) DL4J https://deeplearning4j.org/ S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Execution DL4J SkyMind Cross- platform (JVM) Android Java Java, Scala, Python Y Y - Y Y (Spark)
- 33. • Created & Maintained by Francois Chollet @Google • Release Mar. „2015 • Appliation TensorFlow (http://www.fast.ai/2017/01/03/keras) • Motivation Provide a high-level interface based on deep learning framework like Theano, TensorFlow Easy to use Simple Modular Various Deep-learning framework support Keras https://keras.io/ S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Execution Keras François Chollet Linux, Mac, Windows - Python Python Y(Thean o) N(TF) Y - Y
- 34. • Created by CMU (http://www.cs.cmu.edu/~muli/file/mxnet-learning-sys.pdf) • Maintained by DMLC(Distributed Machine Learning Community) CMU, NYU, NVIDIA, Baidu, Amazon, etc. • Release Oct. „2015 • Application AWS (https://www.infoq.com/news/2016/11/amazon-mxnet-deep-learning) • Motivation Support for Mixed Programming Model: Imperative & Symbolic Support for Portability: Desktops, Clusters, Mobiles, etc. Support for Multiple Languages: C++, R, Python, Matlab, Javascript, etc. MXNet http://mxnet.io/ S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Execution MXNet DMLC Linux, Mac, Windows, Javascript Android, iOS C++ C++, Python, Julia, MATLAB, JavaScript, Go, R, Scala, Perl Y Y - Y Y
- 35. • Created by James Bergstra, Frederic Bastien, etc. (http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf_ Maintained by LISA lab @ Université de Montréal • Release Nov „2010 • Application Keras Lasagne Blocks • Motivation There‟s any. Theano http://deeplearning.net/software/theano/index.html S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Execution Theano Université de Montréal Linux, Mac, Windows - Python Python Y Y - Y
- 36. • Created & Maintained by Ronan Collobert: Research Scientist @ Facebook Clément Farabet: Senior Software Engineer @ Twitter Koray Kavukcuoglu: Research Scientist @ Google DeepMind Soumith Chinatala: Research Engineer @ Facebook • Release Jul. „2014 • Application Facebook, Google, Twitter, Element Inc., etc. • Motivation Unlike Caffe, for research rather than mass market Unlike Theano, easy to use based on imperative model rather than symbolic model Torch http://torch.ch/ S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Execution Torch Ronan, Clément, Koray, Soumith Linux, Mac, Windows Android, iOS C, Lua Lua Y Y Y Y Not officially
- 37. • Created & Maintained by Google Brain • Release Nov. „2015 • Application Google Search Signals (https://www.bloomberg.com/news/articles/2015-10-26/google-turning-its-lucrative- web-search-over-to-ai-machines) Email auto-responder (https://research.googleblog.com/2015/11/computer-respond-to-this- email.html) Photo Search (https://techcrunch.com/2015/11/09/google-open-sources-the-machine-learning- tech-behind-google-photos-search-smart-reply-and-more/#.t38yrr8:fUIZ) • Motivation It‟s Google TensorFlow https://www.tensorflow.org/ S/W Creator Platform Mobile Langua ge Interface OpenMP CUDA OpenCL Multi GPU Parallel Executi on TensorFlow Google Linux, Mac, Windows Android, iOS C++, Python Python, C/C++, Java, Go N Y - Y Y
- 38. Google Tensorflow on github
- 39. * Source: Oriol Vinyals – Research Scientist at Google Brain
- 40. Expressing High-Level ML Computations • Core in C++ • Different front ends for specifying/driving the computation • Python and C++ today, easy to add more * Source: Jeff Dean– Research Scientist at Google Brain
- 41. Hello World on TensorFlow
- 42. Image recognition in Google Map * Source: Oriol Vinyals – Research Scientist at Google Brain
- 43. Deep Learning Hello World == MNIST
- 44. MNIST (predict number of image)
- 45. CNN (convolution neural network) training
- 46. MNIST code
- 50. Multy layer Deep Networks
- 52. Human-Level Face Recognition • Convolutional neural networks based face recognition system is dominant • 99.15% face verification accuracy on LFW dataset in DeepID2 (2014) Beyond human-level recognition Source: Taigman et al. DeepFace: Closing the Gap to Human-Level Performance in Face Verification, CVPR’14
- 53. Deep & Deep Neural network
- 54. ImangeNet Classification Top-5 error
- 55. ImageNet Large Scale Visual Recognition Challenge
- 56. Image Recognition * Source: Oriol Vinyals – Research Scientist at Google Brain
- 57. Object Classification and Detection
- 58. Language Generating * Source: Oriol Vinyals – Research Scientist at Google Brain
- 59. How to the Object recognition ?
- 62. Colorful Image Colorization abL Concatenate (L,ab)Grayscale image: L channel “Free” supervisory signal Semantics? Higher-level abstraction? Ref: Richard Zhang, Phillip Isola, Alexei (Alyosha) Efros
- 63. Inceptions
- 67. Neuro paint
- 68. pix2pix
- 69. Edges2Image [Isola et al. CVPR 2017]
- 70. Image Generate
- 71. 3D Generative Adversarial Network [Wu et al. NIPS 2016]
- 73. Scene Parsing [Farabet et al. ICML 2012, PAMI 2013]
- 74. Scene Parsing [Farabet et al. ICML 2012, PAMI 2013]
- 75. Auto pilot car
- 77. Recurrent Neural Language modeling
- 78. RNN Unfold into DNN over time
- 79. Goolge Natural Translate Machine
- 81. Deep Q-Learning
- 82. AlphaGo Gen1
- 83. How do data science techniques scale with amount of data?
- 85. Thanks you! Q&A