MACHINE LEARNING WITH SAS

Transcription

1 This webinar will be recorded. Please engage, use the Questions function during the presentation! MACHINE LEARNING WITH SAS SAS NORDIC FANS WEBINAR 21. MARCH 2017 Gert Nissen Technical Client Manager Georg Morsing Senior Manager Kaare Brandt Petersen Education & Academic

2 INTRODUCTION GETTING STARTED Agenda Introduction What is Machine Learning? Advanced Models used in Machine Learning Unstructured data Who-am-I Nordic Director, Education & Academic Ph.d. Mathematical Modelling What-about-you?

3 INTRODUCTION WHY IS MACHINE LEARNING HOT? 1 The Game Go machine beats the human world champion 2 Speaking Chinese when you speak English 3 Looking at pictures and understand what you see Team Alpha Go developed an algorithm beating the world champion Lee Sedol in spring Former Kaggle president Jeremy Howard presented this example in his TED Talk: Speach-to-text + translation + text to speach modulated. ImageNET example from Stanford 2014 text formed by algorithm.

4 INTRODUCTION WHAT IS MACHINE LEARNING? Arthur Samuel ( ), USA Pioneer in computer games First self-learning program playing checkers, 1959 [Machine learning is the] field of study that gives computers the ability to learn without being explicitly programmed

5 INTRODUCTION THEORY VS DATA Theory of what happened Function derived from theory Theory based model fitted to data Data of what happened Function which can adapt to just about every data pattern Data driven modelling

6 In machine learning, data speaks louder than theory

7 ADVANCED MODELS A WAY TO DEAL WITH A COMPLEX REALITY

8 APPROACHES BE FLEXIBLE (ADAPTABLE TO MULTIPLE REALITIES)

9 ADVANCED MODELS OVERFITTING AND BALANCE BETWEEN FLEXIBILITY AND DATA POINTS Model complexity (flexibility) Underlying process Complex Overfitting Overfitting Good fit Fitted function Data point Overfitting Good fit Good fit Potentially good models Too simple models Simple Poor fit Poor fit Poor fit Small Large Data Amount

10 ADVANCED MODELS DATA PARTITIONING IS A WAY TO FIND THE BALANCE BETWEEN FLEXIBILITY AND DATA POINTS Data set 40% Training data Find the parameter values (given the flexibility) 30% Validation data Find the right level of flexibility 30% Test data Estimate performance

11 SOME MODELS USED IN MACHINE LEARNING K-Nearest Neighbours Decision Trees Neural Networks Support Vector Machines Flexibility controlled by the number of neighbours included, K. Flexibility controlled by the number of leaf nodes (boxes), which again is controlled by a number of options, such as performance on the validation set, minimum number of observations for splitting, etc. Flexibility typically controlled by the early stopping, that is starting from small weights corresponding to a linear model then letting these grow and change but stopping when the validation error is increasing. Flexibility controlled by the so-called kernel width; a parameter which determines a typical lenght of the data shape.

12 SOME MODELS USED IN MACHINE LEARNING Ensemble Learning Bagging example: Random Forests Boosting example: Adaptive Boosting Flexibility first and foremost controlled by the individual model handles, but the ensemble approach itself (the bagging) is a regularizer, so there may in fact be a need for overall flexibility adjustment this is in some case handled by the number of submodels. Flexibility controlled by the number of trees and the individual flexibility of the trees (the number of leaf-nodes of the trees). Flexibility controlled by the number of boosting steps (T).

13 HOW TO IN SAS MACHINE LEARNING METHODS IN SAS ENTERPRISE MINER

14 HOW TO IN SAS COURSE Machine Learning with SAS 2 day course Hands-on using SAS Enterprise Miner Next: Copenhagen, April Stockholm, May 9-10

15 UNSTRUCTURED DATA AND DEEP LEARNING

16 SOUND SOME SOUND WHAT CAN YOU HEAR? This is what sound looks like for an algorithm 44,1 khz sampling numbers per sec 3 minutes equals 7,938,000 numbers

17 IMAGES THE MNIST DATA SET MNIST data set Handwritten digits Famous ML benchmark data set images 28x28 grayscale = 784 values per image Table rows 785 columns in total (784 input + 1 target)

18 IMAGES THIS IS IMAGES OF HANDWRITTEN DIGITS

19 Images IMAGES TRADITIONAL APPROACH TO IMAGES Image no 21355: 28x28=784 values 1 2 Features Feature extraction N key values to represent the image content

20 DEEP LEARNING WHAT IS DEEP LEARNING? Geoffrey Hinton (1947-*), Godfather of Deep Learning Born in England, Lives in Canada University of Toronto [Deep] learning is a set of methods that allows a machine to be fed with raw data and to automatically discover the representations needed for detection or classification

21 DEEP LEARNING DEEP LEARNING OVER-SIMPLIFIED INTO ONE SLIDE 1 Unsupervised part for finding the optimal representation Input and output must match (as best possible). Then the middle layer act as a compressed representaiton of the full image 2 Supervised learning on the optimal representation = Alive

22 DEEP LEARNING THE CAT PROBLEM Extracting image features of a cat but cats have many forms Brutto list of images Amazon Mechanical Turk: * persons categorizing and sort * img in categories * images of cats Convoluted neural networks (Hinton et al.) 24 millions nodes 140 millions parametes million connections Source: Fei Fei Li, Director of Stanford AI & Vision Lab, TED Talk 2015

23 CONCLUSIONS sas.com

24 HOW TO IN SAS MACHINE LEARNING IN SAS VIYA (AND MANY ADVANCED METHODS COMING UP IN 2017) More info: SAS User Forum in the Nordics, May & June Source:

25 HOW TO IN SAS COURSE Machine Learning with SAS 2 day course Hands-on using SAS Enterprise Miner Next: Copenhagen, April Stockholm, May 9-10

26 SAS COMMUNITY NORDIC Get the presentation from today and continue your learning Join the Nordic SAS Online Community and receive regular activity updates

27 NORDIC WEBINAR SERIES SIGN UP AT Date Title Area January 5.1. News in SAS 9.4 M4 All February 2.2. Efficient SAS programming Programming 7.2. SAS Studio version 3.6 Programming Calculating values and creating parameters in SAS Visual Analytics Visual Analytics March SAS Environment Manager Administration, Data Management Machine Learning with SAS Analytics April News from SAS Global Forum All Graph Builder and Maps with SAS Visual Analytics Visual Analytics May New versions of SAS Visual Analytics Visual Analytics Note: Date and topics are preliminary. Changes can occur.