Deep learning with Keras

1. Location: ODSC 2017 5/4/2017 Deep Learning with Keras 2016 Copyright QuantUniversity LLC. Presented By: Sri Krishnamurthy, CFA, CAP sri@quantuniversity.com www.analyticscertificate.com

2. 2 Slides and Code will be available at: http://www.analyticscertificate.com/ODSC2017

3. - Analytics Advisory services - Custom training & certificate programs - Fintech and Energy Analytics and Infrastructure

4. • Founder of QuantUniversity LLC. and www.analyticscertificate.com • Advisory and Consultancy for Financial Analytics • Prior Experience at MathWorks, Citigroup and Endeca and 25+ financial services and energy customers. • Regular Columnist for the Wilmott Magazine • Author of forthcoming book “Financial Modeling: A case study approach” published by Wiley • Charted Financial Analyst and Certified Analytics Professional • Teaches Analytics in the Babson College MBA program and at Northeastern University, Boston Sri Krishnamurthy Founder and CEO 4

5. 5 Quantitative Analytics and Big Data Analytics Onboarding • Trained more than 1000 students in Quantitative methods, Data Science and Big Data Technologies using MATLAB, Python and R • Launching the Analytics Certificate Program in Summer and a Fintech Certification program in Fall

6. 6 • May 2017 ▫ Sponsoring the CFA Fintech Conference in Boston ▫ QuantUniversity Chicago Meetup  Deep Learning – May 18th - https://www.meetup.com/QuantUniversity- Meetup-Chicago ▫ Deep Learning Workshop – May 30,31st  Chicago & Online : http://www.analyticscertificate.com/DeepLearning • June 2017 ▫ Machine Learning Workshop – June 8th, 9th  New York & Online : http://www.analyticscertificate.com/MachineLearning ▫ Anomaly Detection Workshop – June 18th, 19th  Boston & Online : http://www.analyticscertificate.com/Anomaly Events of Interest

7. Summer 2017: http://www.analyticscertificate.com

8. 8 • Boston • New York • Chicago • Washington DC (Coming soon) QuantUniversity meetups

10. 10 • Unsupervised Algorithms ▫ Given a dataset with variables 𝑥𝑖, build a model that captures the similarities in different observations and assigns them to different buckets => Clustering, etc. ▫ Create a transformed representation of the original data=> PCA Machine Learning Obs1, Obs2,Obs3 etc. Model Obs1- Class 1 Obs2- Class 2 Obs3- Class 1

11. 11 • Supervised Algorithms ▫ Given a set of variables 𝑥𝑖, predict the value of another variable 𝑦 in a given data set such that ▫ If y is numeric => Prediction ▫ If y is categorical => Classification Machine Learning x1,x2,x3… Model F(X) y

12. 12 Start with labeled pairs (Xi, Yi) ( ,“kitten”),( ,“puppy”) …

13. 13 Success: predict new examples ( ,?)

14. 14 https://commons.wikimedia.org/wiki/Neural_network “kitten” “puppy” “has fur?” “pointy ears?” “dangerously cute?”

15. 15

16. 16 http://stackoverflow.com/questions/40537503/deep-neural-networks-precision-for-image-recognition-float-or-double

17. 17 http://stackoverflow.com/questions/40537503/deep-neural-networks-precision-for-image-recognition-float-or-double Weighted sum

18. 18 http://stackoverflow.com/questions/40537503/deep-neural-networks-precision-for-image-recognition-float-or-double Non-linear “activation” function

19. 19 http://stackoverflow.com/questions/40537503/deep-neural-networks-precision-for-image-recognition-float-or-double Learning = “find good weights”

20. 20 http://stackoverflow.com/questions/40537503/deep-neural-networks-precision-for-image-recognition-float-or-double Learning = “find good weights” How? Gradient descent!

21. 22 1. Our labeled datasets were thousands of times too small. 2. Our computers were millions of times too slow. 3. We initialized the weights in a stupid way. 4. We used the wrong type of non-linearity. - Geoff Hinton Neural nets were tried in the 1980s. What changed? https://youtu.be/IcOMKXAw5VA?t=21m29s

22. 23 http://www.rsipvision.com/exploring-deep-learning/

23. 24 http://www.asimovinstitute.org/neural-network-zoo/

24. 25

25. 26 https://research.googleblog.com/2014/09/building-deeper-understanding-of-images.html

26. 27 https://research.googleblog.com/2014/09/building-deeper-understanding-of-images.html

27. 28 Computer-Aided Diagnosis with Deep Learning Architecture: Applications to Breast Lesions in US Images and Pulmonary Nodules in CT Scans http://www.nature.com/articles/srep24454/figures/1

28. 29 Towards End-to-End Speech Recognition with Recurrent Neural Networks http://www.jmlr.org/proceedings/papers/v32/graves14.pdf

29. 30 https://www.technologyreview.com/s/544651/baidus-deep-learning-system-rivals-people-at-speech-recognition/

30. 31 https://research.googleblog.com/2014/11/a-picture-is-worth-thousand-coherent.html

31. 32 http://cs.umd.edu/~miyyer/data/deepqa.pdf

32. 33

33. 34 http://blog.ventureradar.com/2016/03/11/10-hot-startups-using-artificial-intelligence-in-cyber-security/

34. 35 https://www.youtube.com/watch?v=H4V6NZLNu-c

35. 36 https://www.engadget.com/2016/03/12/watch-alphago-vs-lee-sedol-round-3-live-right-now/

36. 37 https://www.youtube.com/watch?v=kMMbW96nMW8

37. 38

38. 39 How is deep learning special? Given (lots of) data, DNNs learn useful input representations. D. Erhan et al. ‘09 http://www.iro.umontreal.ca/~lisa/publications2/index.php/publications/show/247

39. 40

40. 41 Hardware

41. 42 Data http://www.theneweconomy.com/strategy/big-data-is-not-without-its-problems

42. 43 New Approaches http://deeplearning.net/reading-list/

43. 44

44. 45 • Theano is a Python library that allows you to define, optimize, and evaluate mathematical expressions involving multi-dimensional arrays efficiently • Performs efficient symbolic differentiation • Leverages NVIDIA GPU (Claim 140X faster than CPU) • Developed by University of Montreal researchers and is open-source • Works on Windows/Linux/Mac OS • See https://arxiv.org/abs/1605.02688 Theano

45. 46 • GPU vs CPU ▫ Theano Test ▫ See Theano Test.ipyb Demo

46. 47 • Logistic Regression Theano See Theano-Logistic Regression.ipyb

47. 48 MLP

48. 49 Convolutional Neural Networks Convolution

49. 50 Max pooling

50. 51 Convolutional Neural Networks See Theano-Conv-Net.ipynb

51. 52 • Keras is a high-level neural networks library, written in Python and capable of running on top of either TensorFlow or Theano. It was developed with a focus on enabling fast experimentation. • Allows for easy and fast prototyping (through total modularity, minimalism, and extensibility). • Supports both convolutional networks and recurrent networks, as well as combinations of the two. • Supports arbitrary connectivity schemes (including multi-input and multi-output training). • Runs seamlessly on CPU and GPU. Keras

52. 53 • Keras Examples ▫ Testing Keras: See KerasPython.ipynb ▫ Mlp-1 layer ▫ Running Convolutional NN on Keras with a Theano Backend  See Keras-conv-example-mnist.ipynb Demo

53. 54

54. 55 • Motivation1: Autoencoders 1. http://ai.stanford.edu/~quocle/tutorial2.pdf

55. 56 • Goal is to have ෤𝑥 to approximate x • Interesting applications such as ▫ Data compression ▫ Visualization ▫ Pre-train neural networks Autoencoder

56. 57 Demo in Keras1 1. https://blog.keras.io/building-autoencoders-in-keras.html 2. https://keras.io/models/model/

57. Supervised learning Cross-sectional ▫ Observations are independent ▫ Given X1----Xi, predict Y ▫ CNNs

58. Supervised learning Sequential ▫ Sequentially ordered ▫ Given O1---OT, predict OT+1 1 Normal 2 Normal 3 Abnormal 4 Normal 5 Abnormal

59. 60 • Given : X1,X2,X3----XN • Convert the Univariate time series dataset to a cross sectional Dataset Time series modeling in Keras using MLPs X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X Y X1 X2 X2 X3 X3 X4 X4 X5 X5 X6 X6 X7 X7 X8 X8 X9 X9 X10 X10 X11 X11 X12 X12 X13 X13 X14 X14 X15

60. 61 • Monthly data • Computational Intelligence in Forecasting • Source: http://irafm.osu.cz/cif/main.php?c=Static&page=download Sample data 0 200 400 600 800 1000 1200 1400 1600 1800 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 100 103 106

61. 62 • Use 72 for training and 36 for testing • Lookback 1, 10 • Longer the lookback, larger the network Multi-Layer Perceptron Size 8 Size 1

62. 63 Demo Train Score: 1972.20 MSE (44.41 RMSE) Test Score: 3001.77 MSE (54.79 RMSE) Train Score: 2631.49 MSE (51.30 RMSE) Test Score: 4166.64 MSE (64.55 RMSE) Lookback = 1 Lookback = 10

63. 64 • Has 3 types of parameters ▫ W – Hidden weights ▫ U – Hidden to Hidden weights ▫ V – Hidden to Label weights • All W,U,V are shared Recurrent Neural Networks1 1. http://ai.stanford.edu/~quocle/tutorial2.pdf

64. 65 Where can Recurrent Neural Networks be used?1 1. http://karpathy.github.io/2015/05/21/rnn-effectiveness/ 1. Vanilla mode of processing without RNN, from fixed-sized input to fixed-sized output (e.g. image classification). 2. Sequence output (e.g. image captioning takes an image and outputs a sentence of words). 3. Sequence input (e.g. sentiment analysis where a given sentence is classified as expressing positive or negative sentiment). 4. Sequence input and sequence output (e.g. Machine Translation: an RNN reads a sentence in English and then outputs a sentence in French). 5. Synced sequence input and output (e.g. video classification where we wish to label each frame of the video).

65. 66 • Andrej Karpathy’s article ▫ http://karpathy.github.io/2015/05/21/rnn-effectiveness/ • Hand writing generation demo ▫ http://www.cs.toronto.edu/~graves/handwriting.html Sample applications

66. 67 Recurrent Neural Networks • A recurrent neural network can be thought of as multiple copies of the same network, each passing a message to a successor. 1 • Backpropagation(computing gradient wrt all parameters of the network) which is process used to propagate errors and weights needs to be modified for RNNs due to the existence of loops http://colah.github.io/posts/2015-08-Understanding-LSTMs/

67. 68 • BPTT begins by unfolding a recurrent neural network through time as shown in the figure. • Training then proceeds in a manner similar to training a feed- forward neural network with backpropagation, except that the training patterns are visited in sequential order. Back Propagation through time (BPTT)1 1. https://en.wikipedia.org/wiki/Backpropagation_through_time

68. 69 • Backpropagation through time (BPTT) for RNNs is difficult due to a problem known as vanishing/exploding gradient . i.e, the gradient becomes extremely small or large towards the first and end of the network. • This is addressed by LSTM RNNs. Instead of neurons, LSTMs use memory cells 1 Addressing the problem of Vanishing/Exploding gradient http://deeplearning.net/tutorial/lstm.html

69. 70 • Dataset of 25,000 movies reviews from IMDB, labeled by sentiment (positive/negative). • Reviews have been preprocessed, and each review is encoded as a sequence of word indexes (integers). • For convenience, words are indexed by overall frequency in the dataset, so that for instance the integer "3" encodes the 3rd most frequent word in the data. • The 2011 paper (see below) had approximately 88% accuracy • See ▫ https://github.com/fchollet/keras/blob/master/examples/imdb_lstm.py ▫ http://machinelearningmastery.com/sequence-classification-lstm-recurrent-neural- networks-python-keras/ ▫ http://ai.stanford.edu/~amaas/papers/wvSent_acl2011.pdf Demo – IMDB Dataset

70. 71 Network The most frequent 5000 words are chosen and mapped to 32 length vector Sequences are restricted to 500 words; > 500 cut off ; < 500 pad LSTM layer with 100 output dimensions Accuracy: 84.08%

71. 72

72. 73 • Neural Networks are resource intensive ▫ Typically require huge dedicated hardware (RAM, GPUs) • Parameter space huge! – 100s of thousands of parameters ▫ Tuning is important • Architecture choice is important: ▫ See http://www.asimovinstitute.org/neural-network-zoo/ Key takeaways from modeling Deep Neural Networks

73. What is Spark ? • Apache Spark™ is a fast and general engine for large-scale data processing. • Run programs up to 100x faster than Hadoop MapReduce in memory, or 10x faster on disk. Lightning-fast cluster computing

74. Why Spark ? Generality • Combine SQL, streaming, and complex analytics. • Spark powers a stack of high-level tools including: 1. Spark Streaming: processing real-time data streams 2. Spark SQL and DataFrames: support for structured data and relational queries 3. MLlib: built-in machine learning library 4. GraphX: Spark’s new API for graph processing

75. 76 • Investment : Enterprises have significantly invested in Big-Data infrastructure • GPUs – Require specialized hardware – Niche Use-cases • Can enterprises reuse existing infrastructure for deep learning applications? • What use-cases in Deep learning can leverage Apache Spark? Deep Learning + Apache Spark ?

76. 77 • Databricks – Platform for running Spark applications • BigDL – Intel’s library for deep learning on existing data frameworks. • TensorflowOnSpark – Yahoo’s Distributed Deep Learning on Big Data Clusters • The Rest: ▫ SparkNet – AMPLab’s framework for training deep networks in Spark ▫ DeepLearning4J – Uses Data parallism to train on separate neural networks ▫ DeepDist - Lightning-Fast Deep Learning on Spark Via parallel stochastic gradient updates Efforts on using Deep Learning Frameworks with Spark

77. 78 • Deploying trained models to make predictions on data stored in Spark RDDs or Dataframes  Inception model: https://www.tensorflow.org/tutorials/image_recognition  Each prediction requires about 4.8 billion operations  Parallelizing with Spark helps scale operations Databricks https://databricks.com/blog/2016/12/21/deep-learning-on- databricks.html

78. 79 • Distributed model training  Use deep learning libraries like TensorFlow to test different model hyperparameters on each worker  Task parallelism Databricks https://databricks.com/blog/2016/12/21/deep-learning-on- databricks.html

79. 80 • Tensorframes  Experimental TensorFlow binding for Scala and Apache Spark.  TensorFrames (TensorFlow on Spark Dataframes) lets you manipulate Apache Spark's DataFrames with TensorFlow programs.  TensorFrames is available as a Spark package. Databricks https://github.com/databricks/tensorframes

80. 81 • BigDL is an open source, distributed deep learning library for Apache Spark that has feature parity with existing popular deep learning frameworks like Torch and Caffe • BigDL is a standalone Spark package Intel’s BigDL library https://www.oreilly.com/ideas/deep-learning-for-apache-spark

81. 82 • BigDL uses Intel Math Kernel Library, a fast math library for Intel and compatible processors to facilitate multi-threaded programming in each Spark task. • The MKL library facilitates efficiently train larger models across a cluster (using distributed synchronous, mini-batch SGD) • Key Value proposition: ▫ “The typical deep learning pipeline that involves data preprocessing and preparation on a Spark cluster and model training on a server with multiple GPUs, now involves a simple Spark library that runs on the same cluster used for data preparation and storage.” Intel’s BigDL library https://www.oreilly.com/ideas/deep-learning-for-apache-spark

82. 83 • Existing DL frameworks often require setting up separate clusters for deep learning, forcing us to create multiple programs for a machine learning pipeline TensorflowOnSpark, CafeOnSpark – Yahoo’s Distributed Deep Learning https://github.com/yahoo/TensorFlowOnSpark http://yahoohadoop.tumblr.com/post/157196317141/open-sourcing- tensorflowonspark-distributed-deep

83. 84 • TensorFlowOnSpark supports all types of TensorFlow programs, enabling both asynchronous and synchronous training and inferencing. It supports model parallelism and data parallelism. https://github.com/yahoo/TensorFlowOnSpark http://yahoohadoop.tumblr.com/post/157196317141/open-sourcing- tensorflowonspark-distributed-deep TensorflowOnSpark, CafeOnSpark – Yahoo’s Distributed Deep Learning

84. 85 • Developed at UC Berleley’s AMPLab • SparkNet is built on top of Spark and Caffe. • Not much activity in the last year https://github.com/amplab/SparkNet • SparkNet's parallelized stochastic gradient decent (SGD) algorithm requires minimal communication between nodes SparkNet https://arxiv.org/pdf/1511.06051v1.pdf

85. 86 • Deeplearning4j (DL4J) leverages Spark clusters for fast, distributed, in-memory training of DL models that were developed Scala or Java • A centralized DL model iteratively averages the parameters produced by separate neural nets. DeepLearning4J https://deeplearning4j.org/spark.html#how

86. 87 • Leverages Spark and asynchronous SGD to accelerate Deep Learning training from HDFS/Spark data • DeepDist fetches the model from the master and calls gradient(). After computing gradients on the data partitions, gradient updates are sent back the server. On the server, the master model is updated by descent() using the updates from the nodes.. DeepDist http://deepdist.com/

87. 88 • Databricks – Platform for running Spark applications • BigDL – Intel’s library for deep learning on existing data frameworks. • TensorflowOnSpark – Yahoo’s Distributed Deep Learning on Big Data Clusters • The Rest: ▫ SparkNet – AMPLab’s framework for training deep networks in Spark ▫ DeepLearning4J – Uses Data parallism to train on separate neural networks ▫ DeepDist - Lightning-Fast Deep Learning on Spark Via parallel stochastic gradient updates Efforts on using Deep Learning Frameworks with Spark

88. 89 • QuantUniversity has started a new initiative to support students and unemployed professionals interested in fintech and data science roles to attend our workshops for free/reduced cost. • If you or some one you know are interested in attending our workshops for free/significantly discounted price, apply for a scholarship here • If you are want to join us in supporting this initiative through a sponsorship, please contact us. We are on a mission to democratize Analytics education and we seek your support in making it possible! QuantUniversity’s Analytics for a cause Initiative

89. 90 Q&A

90. Thank you! Checkout our programs at: www.analyticscertificate.com Sri Krishnamurthy, CFA, CAP Founder and CEO QuantUniversity LLC. srikrishnamurthy www.QuantUniversity.com Information, data and drawings embodied in this presentation are strictly a property of QuantUniversity LLC. and shall not be distributed or used in any other publication without the prior written consent of QuantUniversity LLC. 91