Machine Learning: Learning structure from data

Machine learning techniques draw on many areas from statistics to theoretical computer science. These algorithms have advanced applications in robotics, health care, finance, and computer games, to name just a few. This talk will serve as a pragmatic introduction to machine learning, understanding of core concepts in a machine learning pipeline, and how one might use it to solve problems in practical settings.

Transcript

1. Machine Learning | Learning structure from data Diana Spencer
2. Hi! ML engineer DIANA SPENCER About me Machine learning Artificial neural networks Computational neuroscience Math Computer science Royalty
3. Agenda A science experiment Using a pipeline as a tool for comparing algorithms. Machine learning 101 The very basics summed up in 15 minutes. 1 2 3 Supervised learning One method of machine learning.
4. Agenda A science experiment Using a pipeline as a tool for comparing algorithms. Machine learning 101 The very basics summed up in 15 minutes. 1 2 3 Supervised learning One type of machine learning method.
5. What is machine learning? Machine learning is the field of study that gives computers the ability to learn without being explicitly programmed. Arthur Samuel, 1959
6. The context The relationship between different AI disciplines MACHINE LEARNING ARTIFICIAL INTELLIGENCE DEEP LEARNING
7. 1952 1957 1970–1985 1989 2017 How did we get here? A brief history of machine learning Checkers Program First supervised learning application Perceptron A neural network learning algorithm AI Winter First practical demo of back propagation CheXNet Rajpurkar et al., CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning, 2017
8. Why use machine learning? To build automated systems with superior human-level intelligence to assist us with difficult problem solving tasks.
9. ƞ Examples The use of machine learning to assist us with problem solving tasks Self-driving cars The Google self-driving car project, Waymo. Cyber security Adaptive learning algorithms that can detect the step-by-step penetration of advanced threats.
10. Why use machine learning? To build automated systems with superior human-level intelligence to assist us with difficult problem solving tasks. To build automated systems to improve the quality of life for people and the environment.
11. Ŵ Examples The use of machine learning to improve the quality of life for people and the environment Google Translate Google’s free service instantly translates words, phrases, and web pages between hundreds of languages. Agriculture Crop health monitoring can provide detailed information about plant health across thousands of acres. Wu et al., Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation, 2016
12. INPUT Data Prediction The three basic steps Learning structure from data OUTPUT BLACK BOX Algorithm
13. But what is data?
14. Examples Data comes in many forms
15. So how does a machine learn from the data? Three main methods
16. Types of machine learning Choosing a method depends on the problem Semi-supervised learning Labels = [???, cat, ???] Supervised learning Labels = [cat, cat, dog] Unsupervised learning Labels = [???, ???, ???]
17. Agenda A science experiment Using a pipeline as a tool for comparing algorithms. Machine learning 101 The very basics summed up in 15 minutes. 1 2 3 Supervised learning One method of machine learning.
18. Supervised learning One type of machine learning SUPERVISED LEARNING SUPPORT VECTOR MACHINE CONVOLUTIONAL NEURAL NETWORKS LINEAR REGRESSION REGRESSION CLASSIFICATION
19. Linear regression One type of supervised learning
20. Supervised learning One type of machine learning SUPERVISED LEARNING SUPPORT VECTOR MACHINE CONVOLUTIONAL NEURAL NETWORKS LINEAR REGRESSION REGRESSION CLASSIFICATION
21. “AND” logic gate Mapping inputs to targets to create samples
22. . . . 40 “AND” logic gate Creating more samples
23. “AND” logic gate Using the samples to train the algorithm A B Target ALGORITHM
24. “AND” logic gate Measuring the error of a prediction ERROR = (TARGET – OUTPUT) ALGORITHM
25. “AND” logic gate The decision boundary line
26. “XOR” logic gate Mapping inputs to targets to create samples
27. “XOR” logic gate Creating more samples
28. “XOR” logic gate The decision boundary line
29. “XOR” logic gate The decision boundary line
30. Supervised learning One type of machine learning SUPERVISED LEARNING SUPPORT VECTOR MACHINE CONVOLUTIONAL NEURAL NETWORKS LINEAR REGRESSION REGRESSION CLASSIFICATION
31. Agenda A science experiment Using a pipeline as a tool for comparing algorithms. Machine learning 101 The very basics summed up in 20 minutes. 1 2 3 Supervised learning One method for machine learning.
32. Comparing two algorithms Using a pipeline as a tool for determining which algorithm has the lowest error rate Non-deep machine learning algorithm Deep machine learning algorithm SUPPORT VECTOR MACHINE CONVOLUTIONAL NEURAL NETWORKS
33. A machine learning pipeline Quick overview COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY Evaluate model Hyper- parameter tuning
34. Training a Support Vector Machine Non-deep machine learning algorithm
35. 16.04 LTS COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY Dogs vs. Cats dataset from kaggle.com
36. Reformatting images to desired inputs COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
37. Reduce the number of image pixels 150 pixels 150 pixels 98 150 x 150 x 3 = 67,500 INPUT VALUES! RGB COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
38. Convert from RGB to gray scale 150 pixels x 150 pixels = 51,529 INPUT VALUES COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
39. Applying PCA COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
40. TRAINING COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY Split data into three groups TRAINING PHASE STEP 01 ENTIRE DATASET TEST VALIDATION
41. Let’s train our first classifier COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
42. Results COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
43. Training a Convolutional Neural Network Deep, machine learning algorithm
44. A simple neural network concept Mapping a biological neuron to an artificial neuron Cell body Dendrites Axon Axon Terminals Axon hillock
45. AlexNet: A type of convolutional neural network Krizhevsky et al., ImageNet Classification with Deep Convolutional Neural Networks, 2012 INPUT OUTPUT
46. Convolutional Neural Network in code
47. 16.04 LTS COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY Dogs vs. Cats dataset from kaggle.com
48. Reading and resizing image using Tensorflow COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
49. Training our model COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
50. Evaluating our model COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
51. Evaluating our model DOG 99.0% CAT 98.0% DOG 92.0% COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
52. Accuracy and loss curves Results COLLECT DATA DATA PRE-PROCESSING TRAIN CLASSIFIER DEPLOY
53. http://www.wildml.com/2015/11/understanding-convolutional-neural-networks-for-nlp/ http://colah.github.io/posts/2015-08-Understanding-LSTMs/ https://deeplearning4j.org/generative-adversarial-network A zoo of models Convolutional neural network Recurrent neural network Generative Adversarial network
54. Real-world business cases Beyond offering interesting AI products Predict demand Predict trends Optimize warehouse and logistics Predict costumer orders Diagnose patients remotely Make quicker diagnoses Customize treatments Retail Health Bughin et al., McKinsey Global Institute Discussion Paper: Artificial Intelligence The Next Digital Frontier? 2017
55. Real-world business cases Core problem in the field: Bias and Inequality https://www.nytimes.com/2016/06/26/opinion/sunday/artificial-intelligences-white-guy-problem.html?_r=1
56. INPUT Data Prediction OUTPUT BLACK BOX Algorithm Predictive algorithms are only as good as the data they are trained on
57. How to cure bias? It’s not a simple solution. Companies who are building high stakes decision-making systems • Building an interdiscipary group to test how your systems work against different categories • Build groups and collaborations that tackle ethical questions • Question who your system will benefit and who might be harmed
58. A well-thought out pipeline is an essential tool …particularly for evaluating and comparing models. Deep learning is a powerful method for tackling difficult problems Sometimes it requires less pre-processing. Machine learning requires skills It is a science and an art. 1 2 3 What did we find out? Summary
59. Thank you! diana.spencer@webstep.no medium.com/@diana.spencer
60. Questions?