May Masoud SAS Canada

Transcription

1 May Masoud SAS Canada

2 #ROAD2AI

3 #ROAD2AI

4 Artificial Intelligence is the science of training systems to emulate human tasks through learning and automation.

5 General Intelligence Robotics Advanced Automation #ROAD2AI

6 #ROAD2AI

7 Strong AI Full AI General Intelligence Artificial General Intelligence #ROAD2AI

8 #ROAD2AI

9 Robotics #ROAD2AI

10 #ROAD2AI

11 AI for Business Fraud Analysis Image Recognition Customer Experience Pragmatic AI Advanced Automation Intelligent Automation Practical AI #ROAD2AI

12 General Intelligence Robotics Advanced Automation #ROAD2AI

13 Advanced Automation #ROAD2AI

14 AI is a Spectrum Rules-based systems Simplest form of automation, the execution of rules Predictive Analytics Predict, advise, influence, recommend Machine Learning Advanced analytic algorithms create insight with more automation Text Processing The addition of unstructured text Automation Deep Learning Self learning algorithms that deliver even more insight and automation Natural Language Understanding Both natural language ingestion and generation. Computer Vision The addition of images and video Robotics Automate repetitive functions and processes

15 What Does AI Do? Find The Outliers See The Future Learn Patterns Recognize Objects Detect anomalous or unexpected behaviors. Predict likely outcomes. Find unknown patterns and relationships. Categorize or catalog like people or things. #ROAD2AI

16 4 Types of Algorithms Teaching by example. The machine uses examples to determine logic. SUPERVISED REINFORCED Learning a game. The rules of the game are provided. The machine takes actions and uses results to learn how to win. Similar to supervised learning. Targets are only provided for a subset of the data. No answer key is provided. Conclusions are unconstrained. Draws inferences and conclusions based solely on analyzing input data. SEMI-SUPERVISED UNSUPERVISED

17 Supervised Learning Teaching by example. Similar to providing an answer key and asking the student (machine) to show their work. The machine uses labeled examples to determine appropriate logic or algorithm.

18 Supervised Learning: Common Techniques Bayesian Statistics Decision Trees Gradient Boosting Neural Networks Random Forests Regression Analysis Support Vector Machines [SVM]

19 Supervised Learning: Practical Applications PERSONALIZATION INTERACTION FRAUD DETECTION NETWORK OUTAGE RISK MANAGEMENT CUSTOMER CHURN

20 Unsupervised Learning No answer key is provided. Conclusions are unconstrained. Modeled on how humans observe the world. Draws inferences and conclusions based solely on analyzing input data.

21 Unsupervised Learning: Common Techniques Affinity Analysis Clustering Clustering: K-Means Nearest-Neighbor Mapping Self-Organizing Maps Singular Value Decomposition

22 Unsupervised Learning: Practical Applications ANOMALY/INTRUSION DETECTION IDENTIFYING LIKE THINGS MARKET BASKET ANALYSIS

23 Semi-Supervised Learning Similar to supervised learning. Q&A are only provided for a subset of the data. Used when there is too much data or subtle variations in data to allow for a comprehensive set of examples.

24 Semi-Supervised Learning: Common Techniques Bayesian Statistics Decision Trees Forecasting Neural Networks Random Forests Regression Analysis Support Vector Machines [SVM]

25 Semi-Supervised Learning : Practical Applications SPEECH RECOGNITION IMAGE RECOGNITION CLASSIFICATION WEB PAGE CLASSIFICATION

26 Reinforcement Learning Similar to teaching someone to play a game. The rules of the game are provided: allowed actions, rules and potential end states. The machine takes different actions and observes results to learn how to achieve an optimal or optimized result.

27 Reinforcement Learning: Common Techniques Artificial Neural Networks (ANN) Learning Automata Markov Decision Process (MDP) Q-Learning

28 Reinforcement Learning: Practical Applications IMAGE RECOGNITION WITH ROBOTICS NAVIGATION GAMING

29 Impact of Better Algorithms Analytic Accuracy Traditional Methods Machine Learning Lost opportunity Opportunity captured #ROAD2AI

30 What do you expect from AI? Better Results Auto Magic Lower Cost Human Interfaces Less assumptions Better math More data including speech, text and images Throw data at it Magic Algorithms Automate the processes Machines do the work More Productivity Better User Experience Less human intervention #ROAD2AI

31 #ROAD2AI

32 Navigate Through The Hype Explain Manage Scale #ROAD2AI

33 Navigate Through The Hype Explain Manage Scale #ROAD2AI

34 Question Technique Usage What are the top inputs? Decision Tree Surrogate Model Find the main drivers for an output How do the drivers work? Partial Dependence (PD) Gives marginal effect of selected input variable (or multiple variables on target) Individual Conditional Expectation (ICE) Can help to identify subgroups (additive) effects and interactions What is the explanation for a particular predication? Local Interpretable Model-Agnostic Explanation (LIME) Gives explanations for individuals prediction from a classifier #ROAD2AI

35 #ROAD2AI

36 Navigate Through The Hype Explain Manage Scale #ROAD2AI

37 Model Management & Governance Govern expanding model inventory Reduce rising validation and deployment costs Automate deployment Open APIs #ROAD2AI

38 Text analytics Data mining + Machine learning Open source Visual Statistics SAS Studio #ROAD2AI

39 Navigate Through The Hype Explain Manage Scale #ROAD2AI

40 Diverse Interfaces Programming in Language of Choice #ROAD2AI

41 Diverse Interfaces GUI/Point and Click #ROAD2AI

42 SAS Best Practice User created templates #ROAD2AI

43 Hyper-Parameter Auto-Tuning Modeling Algorithm Default Solver Genetic Algorithm + Validation Data Objective Average Validation Results: 10 random partitions, 10 datasets

44 Fault Tolerance Parallelized Algorithms High Performance High Availability Deployment Choices #ROAD2AI

45 Deployment of AI

46 Customer Experience Fraud Analysis Image Recognition #ROAD2AI

47 Performance Assessments

48 Cancer Detection

49 Image Recognition Medical Image Recognition on CT Scans Colon cancer are second most common cancers world wide Introduce a better selection of patients using Liver CT-scans Deep Learning in Genomic and Medical Image Data Predict outcome in patients with liver metastases from colon and rectal cancer

50