Machine Learning, Deep Learning and Data Analysis Introduction
Download as pptx, pdf7 likes6,298 views
The document provides an introduction and overview of machine learning, deep learning, and data analysis. It discusses key concepts like supervised and unsupervised learning. It also summarizes the speaker's experience taking online courses and studying resources to learn machine learning techniques. Examples of commonly used machine learning algorithms and neural network architectures are briefly outlined.
More Related Content
What's hot (20)
Similar to Machine Learning, Deep Learning and Data Analysis Introduction (20)
More from Te-Yen Liu (6)
Recently uploaded (20)
Machine Learning, Deep Learning and Data Analysis Introduction
- 1. 工業技術研究院機密資料 禁止複製、轉載、外流 ITRI CONFIDENTIAL DOCUMENT DO NOT COPY OR DISTRIBUTE Machine Learning, Deep Learning and Data Analysis 簡介 劉得彥 [email protected]
- 2. Outline 2 • Overview of ML, DL and Data Analysis • What is Machine Learning ▪ Take a Look At Linear Regression ▪ Other ML Algorithms at a Glance ▪ What is Neural Network? • What is Deep Learning? • Deep Learning using TensorFlow • Data Analysis ▪ Case 1, 2 and 3 ▪ Multivariate Analysis
- 3. My Experience for Machine Learning 3 • 學習過程走了一些冤望路!! ▪ Hope giving you an experience and guideline • Take courses: ▪ Coursera: Machine Learning ( Got Certificate ) ▪ Udemy: Data Science: Deep Learning in Python ( ongoing) • Study on-line resources: ▪ Youtube、ML/DL tutorials … and so on ▪ https://morvanzhou.github.io/ ▪ http://bangqu.com/gpu/blog ▪ http://www.jiqizhixin.com/insights • Get you hands dirty ▪ Python programming a. TensorFlow Deep Learning Library b. Scikit-Learn Library c. Numby, Pandas, matplotlib, …
- 4. From AI to Deep Learning 4 • 推薦觀賞: 人工智能 极客公园 2017 年 大会演讲 ▪ Google首席科学家李飛飛 https://www.youtube.com/watch?v=uZ-7DVzRCy8 https://blogs.nvidia.com.tw/2016/07/whats-difference-artificial-intelligence-machine-learning-deep-learning-ai/ CPU/ GPU Big Data Algorithms Breakthrough
- 5. ML, DL and Data Analysis 5 • Visually Linking • What we focus today https://whatsthebigdata.com/2016/10/17/visually-linking-ai-machine-learning-deep-learning-big-data-and-data-science/ 迷思?? Data Analysis
- 7. Machine Learning definition 7 • Arthur Samuel (1959). Machine Learning: Field of study that gives computers the ability to learn without being explicitly programmed. • Tom Mitchell (1998) Well-posed Learning Problem: A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P, improves with experience E.
- 8. Machine Learning definition 8 • Suppose your email program watches which emails you do or do not mark as spam, and based on that learns how to better filter spam. What is the task T in this setting? ▪ Classifying emails as spam or not spam. (T) ▪ Watching you label emails as spam or not spam. (E) ▪ The number (or fraction) of emails correctly classified as spam/not spam. (P) ▪ None of the above—this is not a machine learning problem
- 9. What is Machine Learning ? • Without writing any custom code specific to the problem • Feed data to the generic algorithm • It builds its own logic
- 10. Two styles of Machine Learning • Supervised Learning 監督式學習 • Unsupervised Learning 非監督式學習 Use the logic to predict the sales price figure out if there is a pattern or grouping or something Features Label
- 11. What are machine learning algorithms? • Regression Algorithms ▪ Linear Regression ▪ Logistic Regression ▪ LASSO • Decision Tree Algorithms ▪ Classification and Regression Tree (CART) ▪ Iterative Dichotomiser 3 (ID3) ▪ C4.5 and C5.0 (different versions of a powerful approach) • Bayesian Algorithms ▪ Naive Bayes • Clustering Algorithms (unsupervised) ▪ k-Means • Support Vector Machines • Principal Component Analysis • Anomaly Detection • Recommender Systems • Artificial Neural Network Algorithms
- 12. LET’S TAKE A LOOK AT LINEAR REGRESSION 12
- 13. Linear Regression 13 • The Hypothesis Function • Cost Function • Gradient Descent for Multiple Variables https://www.coursera.org/learn/machine-learning/
- 14. Gradient Descent 14 • How to choose learning α https://www.coursera.org/learn/machine-learning/
- 15. Gradient Descent 15 • Convergence of gradient descent with an appropriate learning rate α Cost Function https://www.coursera.org/learn/machine-learning/
- 16. Linear Regression 16 • Training data with linear regression fit https://www.coursera.org/learn/machine-learning/
- 17. OTHER ML ALGORITHMS AT A GLANCE 17
- 18. Logistic Regression 18 • Training data with decision boundary linear decision boundary no linear decision boundary https://www.coursera.org/learn/machine-learning/
- 19. Support Vector Machines 19 • The difference between the kernels in SVM ▪ Linear ▪ Polynomial ▪ Gaussian (RBF) ▪ Sigmoid • SVM (Gaussian Kernel) Decision Boundary ▪ Choose gamma ( auto ) Gaussian (RBF) Non-linear decision boundary https://www.coursera.org/learn/machine-learning/
- 20. K-Means 20 • The original 128x128 image with 24-bit color (three 8-bit ) • using K-means (K=16) to use find the 16 colors that best group (cluster) the pixels in the 3-dimensional RGB space. • K=3 and computing centroid means Iteratively https://www.coursera.org/learn/machine-learning/ 非監督式學習
- 21. Principal Component Analysis • An example to deal with image dimension reduction and proximate recovery. Faces Dataset Recovered faces Principal components 非監督式學習
- 22. WHAT IS NEURAL NETWORK (We will review the previous concepts a little bit) 22
- 23. ML -- write that program by ourselves • To estimate the price of a house ▪ If we could just figure out the perfect weights to use that work for every house, our function could predict house prices! ▪ How to do that with ML? def estimate_house_sales_price(num_of_bedrooms, sqft, neighborhood): price = 0 # a little pinch of this price += num_of_bedrooms * .841231951398213 price += sqft * 1231.1231231 price += neighborhood * 2.3242341421 price += 201.23432095 return price
- 24. ML -- write that program by ourselves • Step 1 ▪ Initialize weights to 1.0 • Step 2 ▪ See the difference and how far off the function is at guessing the correct price • Step 3 ▪ Repeat Step 2 over and over with every single possible combination of weights.
- 25. ML -- What about that whole “try every number” bit in Step 3? θ is what represents your current weights. J(θ) means the ‘cost for your current weights’. • Clever ways to quickly find good values for those weights without having to try very many. • If we graph this cost equation for all possible values of our weights for number_of_bedrooms and sqft Gradient Descent
- 26. Making Smarter Guesses 26 • We ended up with this simple estimation function def estimate_house_sales_price(num_of_bedrooms, sqft, neighborhood): price = 0 # a little pinch of this price += num_of_bedrooms * 0.123 price += sqft * 0.41 price += neighborhood * 0.57 return price a linear relationship with the input
- 27. If there is more complicated situation? ▪ Different of weights for the different house sizes
- 28. What is a Neural Network • Now we have four different price estimates. • Let’s combine those four price estimates into one final estimate. neurons This is a neural network
- 29. What is a Neural Network • Human Brains http://www.slideshare.net/tw_dsconf/ss-62245351
- 30. What is a Neural Network • Different connections lead to different structured network. http://www.slideshare.net/tw_dsconf/ss-62245351
- 33. Fully connected feedforward network • Matrix Operation http://www.slideshare.net/tw_dsconf/ss-62245351
- 34. Output Layer • Softmax layer as the output layer
- 35. Neural Network Playground • http://playground.tensorflow.org/
- 37. What is Deep Learning? 37 • Deep Learning is Large Neural Networks • Deep Learning attracts lots of attention http://static.googleusercontent.com/media/research.google.com/en//people/jeff/BayLearn2015.pdf
- 38. Why Deep Learning? • The more data, the more performance. • Game Changer ▪ DL accuracy/performance is more than 99%
- 39. Deep Learning Models 39 • Convolutional Neural Network ▪ Inception-V3 • Recurrent Neural Network ▪ LSTM • Auto-encoder • Reinforcement Learning ▪ Q-Learning ▪ Policy Gradient • Wide and Deep Learning ▪ Recommender system
- 40. Deep Learning is not so simple • Backpropagation ▪ an efficient way to compute Gradient Descent • Overfitting • Choosing Loss function ▪ Square Error, Cross Entropy, and so on… • Mini-Batch • Too deep ( many hidden layers ) ▪ ReLU, MaxOut, … • Learning Rates • Momentum ▪ Adam ( optimizer ) • Weight Decay • Dropout
- 41. Backpropagation 41 • A common method of training artificial neural networks and used in conjunction with an optimization method such as gradient descent.
- 42. Underfitting and Overfitting • Bias-Variance Tradeoff
- 43. Convolutional Neural Network (CNN) • Why CNN is for image? ▪ The first layer of fully connected network would be very large
- 44. The solution is Convolution
- 45. The solution is Convolution
- 46. Convolutional Neural Network (CNN) Adding Even More Steps
- 47. Convolutional Neural Network (CNN)
- 49. Linear Regression in TensorFlow 49 X_data array([ 0. , 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. , 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2. , 2.1, 2.2, 2.3, 2.4, 2.5, 2.6, … Y_data array([ 0. , 0.29999667, 0.59997333, 0.89991 , 1.19978668, 1.49958339, 1.79928013, 2.09885695, 2.39829388, 2.69757098, 2.99666833, 3.29556602, …
- 50. Linear Regression in TensorFlow 50
- 51. Linear Regression in TensorFlow 51
- 52. MNIST 52 • 手寫數字辨識 • MNIST dataset ▪ 55000 samples (50000 for training, 5000 for testing) ▪ For each sample, it has X, y parts. ▪ X are the image with 28*28 pixels in 8 bit gray scale ▪ Y is the label answer: 0, 1, 2, …, 9
- 53. MNIST 53 • X, y can be represented as follows
- 54. MNIST 54 • If you want to get the accuracy more than 99%, check it out: • https://gotocon.com/dl/goto-london-2016/slides/MartinGorner_TensorflowAndDeepLearningWithoutAPhD.pdf 92%
- 55. Image Recognition and Retraining • Inception-v3 model is ready and made by Google ▪ it took Google researchers two weeks to build on a desktop with eight NVidia Tesla K40s. ▪ It can recognize > 1000 categories • Retraining ▪ To prepare the new images and categories ▪ Do training and testing
- 56. Plate Number Recognition • There is an example using UK’s Plate Number and Character to train TensorFlow CNN model • Take 3 days with GPU Card (GTX 750 TI) http://matthewearl.github.io/2016/05/06/cnn-anpr/
- 57. Technical breakthrough for Deep-ANPR http://matthewearl.github.io/2016/05/06/cnn-anpr/
- 58. Autoencoder • Enccode the input data (MNIST data) and then decode it back • It is similar to PCA autoencoder original
- 59. LSTM (RNN) • It is a special kind of RNN, capable of learning long-term dependencies LSTM Training with
- 60. Recurrent Neural Network 60 RNN Model training output Play this map with Super Mario Maker https://medium.com/@ageitgey/machine-learning-is-fun-part-2-a26a10b68df3
- 62. Data Analysis 62 • The steps to do data analysis ▪ Data Collection a. From CSV files, database, … and so on. ▪ Data Preprocessing ( very very very important ) a. Regularization, Normalization, … b. Table Join, ▪ Feature Extraction a. Reduce the dimensions …. ▪ Feature Selection a. Select the important features ▪ Machine Learning / Deep Learning a. To train the model b. To do the Prediction and Classification by the trained model c. Apply or implement to system • But, still needs: ▪ domain experts involved!! ▪ Studying related papers and researches
- 63. Analysis Tools and Libraries 63 • Open Sources(Python) ▪ Machine Learning a. SciKit-Learn b. NumPy c. Matplotlib d. Pandas ▪ Deep Learning a. TensorFlow b. Keras ▪ Hadoop & Spark • Commercial Software ( rare to use…) ▪ PolyAnalyst 6.5 ▪ SAS
- 64. Data Analysis 64 • In my experience with data analysis, I belong to a “Rookie”… • 製造資料科學：從預測性思維到處方性決策 ▪ http://www.slideshare.net/tw_dsconf/ss-71780267
- 65. Reference • Machine Learning is Fun! – Medium • Machine Learning is Fun! Part 2 – Medium • Machine Learning is Fun! Part 3: Deep Learning and ... - Medium • Deep Learning Tutorial • FIRST CONTACT WITH TENSORFLOW • https://ireneli.eu/2016/02/03/deep-learning-05-talk-about-convolutional-neural- network%EF%BC%88cnn%EF%BC%89/ • http://www.slideshare.net/tw_dsconf/ss-71780267 • https://morvanzhou.github.io/tutorials/python-basic/ • https://media.readthedocs.org/pdf/python-for-multivariate-analysis/latest/python-for- multivariate-analysis.pdf • http://blog.topspeedsnail.com/ • http://www.leiphone.com/news/201702/vJpJqREn7EyoAd09.html • Python 之機器學習套件 scikit-learn ▪ https://machine-learning-python.kspax.io/ ▪ version >= 0.17