Machine Learning AND Deep Learning for OpenPOWER
- 1. OpenPOWER Webinar Series Machine Learning and Deep Learning 101 Clarisse Taaffe Hedglin [email protected] Executive AI Architect IBM Systems
- 2. 2 Please note IBM’s statements regarding its plans, directions, and intent are subject to change or withdrawal without notice and at IBM’s sole discretion. Information regarding potential future products is intended to outline our general product direction and it should not be relied on in making a purchasing decision. The information mentioned regarding potential future products is not a commitment, promise, or legal obligation to deliver any material, code or functionality. Information about potential future products may not be incorporated into any contract. The development, release, and timing of any future features or functionality described for our products remains at our sole discretion. Performance is based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput or performance that any user will experience will vary depending upon many factors, including considerations such as the amount of multiprogramming in the user’s job stream, the I/O configuration, the storage configuration, and the workload processed. Therefore, no assurance can be given that an individual user will achieve results similar to those stated here.
- 3. Session Objectives Introducing Machine Learning and Deep Learning (ML/DL) in the context of AI and analytics Understanding the iterative nature of the workflow Getting an overview of different ML/DL algorithms Interpreting models and outcomes
- 4. • No governance • No collaboration • Limited complexity How Customers Do Data Analytics Traditionally Spreadsheets • Broad rules and categories • Not dynamic Business Rules • Hard to maintain • Pre-set rules and approaches Homegrown Applications • Limited use of analytics • Hard coded models that do not apply to unique needs • Slow response Other Applications
- 6. Machine Learning Definition 6 “…an application of artificial intelligence (AI) that provides systems the ability to automatically learn and improve from experience without being explicitly programmed.”
- 7. 7 From Data to Actions 010101010101010111100010011001010111 0000000000010101010100000000000 111101011 11000 000000000000 111111 010101 101010 10101010100 Prescriptive What should we do ? Descriptive What Has Happened? Cognitive Learn Dynamically Predictive What Will Happen? ACTIONDATA HUMAN INPUTS < < > < > >c c c c >
- 8. Machine Learning Flow Credit card transaction Loan application MRI image House data Fraudulent vs. legitimate Approve vs. reject Tumor benign vs. malignant House appraisal value Mathematical Function Not a memorization caching system Representing pattern by a mathematical function Machine learning is just a bunch of math prediction
- 9. Data – Estimate House Price Every column except last is a feature Last column is a label This is a labeled data set Sq Ft Bedroom Bathroom Price 2000 3 2 $350,000 1500 2 2 $280,000 2200 3 3 $400,000 … … … …
- 10. With labeled data, it is called supervised machine learning. What if we don’t have labeled data? –It’s unsupervised learning –Objective is forming clusters based on data Machine Learning Categories Customer Revenue Customer Profit # of Online Purchases # of Store Purchases …
- 11. Unstructured, Landing, Exploration and Archive Operational Data Real-time Data Processing & Analytics Transaction and application data Machine, sensor data Enterprise content Image, geospatial, video Social data Third-party data Information Integration & Governance Data is Prerequisite to AI Risk, Fraud Chat bots, personal assistants Supply Chain Optimization Dynamic Pricing, Recommenders Behavior Modeling Vision, Autonomous Systems
- 12. 12 Iterative Process of Machine Learning Define Problem Prepare Data Train Model Evaluate Model Fine-Tune Model Deploy Model
- 13. Prior to performing machine learning, identify foundational knowledge of the problem you are trying to solve. Articulate use case, and identify problem What type of supporting data is available? How large is the data? Identify response time/throughput characteristics Business Understanding
- 14. Predict a Future Event Segment Data / Detect Anomalies Determine optimal quantity, price, resource allocation, or best action Understand Past Activity Discover Insights in Content (text, images, video) Interact in Natural Language Forecast and Budget based on past activity Supervised Unsupervised Predictive: What will happen? Prescriptive: What should we do? Descriptive: What happened? Planning: What is our Plan? NLPDeep Learning Supervised Common Patterns of Analytics Business Problems Solving business problems with Data and AI will utilize a combination of these analytics patterns
- 15. 15 Learning to Map Input to Output Input Output Application Customer behaviour Responder (1/0) Target marketing Banking transactions Fraud (1/0) Fraud Detection Call Data Record Churn (1/0) Customer retention Image Object/Caption (1,…1000) Object Detection Audio Text Transcript Speech recognition Arabic English Machine translation
- 16. Visualization can give us an intuition about the relationships between the data Can be used to find clusters and patterns in data Can be used to understand distributions in data (Univariate statistics) Correlations and Bivariate statistics Many graphical packages available Prepare Data - Visualization
- 17. Feature engineering Feature Selection What to do with missing values How to handle non numeric types Feature A Feature B Feature C Derived Feature A Derived Feature B Target / Label 0 Sentence 1 f1(A,B) f2(A,B) category1 NaN Sentence 2 category2 It has been stated, 80% of a data scientists time is spent in data preparation … here’s why (even after data is identified/obtained) More derived features (binary encoded) Prepare Data What is the best representation of sample data for more accurate prediction?
- 18. An input image of size 256 x 256 contains 65,536 pixels Each pixel is a feature in the feature vector Highly computationally intensive due to large number of features Image Data Representation x11 x12 … x1, 65536 x21 x22 … x2, 65536 x31 x32 … x3, 65536 … … … … xn1 xn2 xn3 xn, 65536 x1 x2 x3
- 19. Train / Test Data Split Break input data All data Training Cross Validation Test Random Sample 1. Training data used to build models 2. Cross validation set used to evaluate similar versions of models 3. Test set is used to inference results to evaluate the quality of the model
- 20. Model Objective Use training data to derive f(x) so that:Cost (Actual - f(x)) Minimize ( Actual - Prediction ) Every computational iteration analyzes the entire training data set Process of minimizing this cost function is called training
- 21. 21 Select Machine Learning Algorithms Supervised Learning Logistic Regression Decision trees Random forests Neural Network (Deep Learning) Bayesian Techniques Support Vector Machines Ensemble Methods Markov Logic Network Unsupervised Learning K-Means Hierarchical Clustering Anomaly Detection Density-based methods Principal Component Analysis
- 22. Logistic Regression Classification system Medical image - tumor benign or malignant Credit card transaction - normal or fraudulent Customer churning - yes or no Email - normal or spam Language identification - English vs. French vs. Spanishx1 x2 f(x) Map input to one of the output categories
- 23. Linear Regression Model output is in continuous real number space Could be multi-dimensional f(x)
- 24. K-Means x1 Randomly choose 3 data points as centroids For each data point, assign them to one of the groups based on distance from the centroids Recompute centroids in each group Reassign each data point Repeat until convergence Identifies patterns and clusters
- 25. Deep Learning Convolutional Neural Network (CNN)
- 26. Recurrent Neural Network (RNN) Deep Learning
- 28. 28 Deep learning (DL) framework invented and open sourced by Google Based on the notion of tensors which are multi dimensional arrays of numbers Implements a number of functions that are common to all deep learning workflows (optimizers, back propagation,etc) Programming Model : User defines neural network as a graph, and then user “feeds” data to the network to either train or perform inference Most widely adopted platform in the DL universe as of today One of the best documented frameworks TensorFlow – a quick review
- 29. IBM Internal Use Only PYTORCH – a quick review 2 Facebook’s framework for research • Cousin of LUA based Torch framework, but was rewritten to be tailored to Python frontend • Gaining popularity quickly for its ease of use in R&D • Supports dynamic computation graphs • Based on Python with Numpy compatibility • Multi-GPU • Easy to use, and supports standard debug tools
- 30. 30 Transfer Learning Useful because you can use pretrained networks that might have taken weeks to train Useful because early layers are trained to distinguish coarser features Typically final layer is removed for a new problem and network is retrained using new data Applying learning acquired in one domain to a problem of a similar domain
- 31. IBM Internal Use Only Definitions: Scratch, Finetune, Feature Extract 31 • Scratch: training all parameters yourself starting from Random • Finetune: training many or most parameters yourself but retain some prior weights • Feature Extract = Pre- Trained: Only train the vary last classification layers
- 32. 32 Agent w/ DNN Environment reward Rt state St St+1 Rt+1 action At Reinforcement Learning Maximizes long term reward • Maximize a dimension over many steps • Correlates immediate actions with delayed benefits Requires a lot of data to learn; often used with a simulator Use Cases: Robotics, Stock pricing, Game adversaries
- 33. 33 Machine Learning Algorithms Input Feature Extraction Features Classification Output Machine Learning Input Output Feature Extraction & Classification Deep Neural Network Deep Learning Model Interpretation
- 34. 34 Humans 5% Error26% Errors Machine Learning Based 3% Errors Deep Learning Based 20162011 Evaluation of Model Accuracy
- 35. 35
- 36. Top 5 Error Rate
- 37. 37 Model Goal – Does it Help my Business ? Targetsin% Optimal Random Model 5010 100 100 Population in % (ranked in descending order of scores) 40 90 100 10
- 38. 38 Model Output Should be Easy to Consume In a Call Center In a Mobile App Deployment is the idea of making the insights available to application developers, consumers, and business users
- 39. Packaged Goods Quality Inspection with Visual Insights Manufacturer Inventory of Manufactured Goods Retailer Shelf Inventory CheckPackage CheckInventory Management Inspection Points Quality Check Images Videos Visual Inspector Industrial Cameras ❶Upload images/videos ❷Train models ❸Inference IBM Maximo Visual Insights Supply Chain Solution Consumer Packaged Goods *No Data Scientist required QUALITY INSPECTION
- 40. Actionable Decisions Image Sensor Database Text Data Fusion Sources Fact Tables AI & Analytics Language Analytics Classification Detection Time Series Descriptive Statistics OpenPOWER based NoviLens AI Appliance Under the Hood – End to End Workflow 40© 2020 NoviSystems Corporation Questions ? ** TechData IBM Reseller in Research Triangle Park *Patents pending on the methods employed by NoviLens DATA FUSION
- 41. AI demands purpose-built infrastructure throughout the journey •Data preparation •Model development environment •Runtime environment •Train, deploy and manage models •Business KPI, production metrics •Explainability and fairness
- 42. Enterprise Data Pipeline with IBM Spectrum Storage 42 Insights Out Trained Model Inference Data In Transient Storage SDS/Cloud Global Ingest Throughput-oriented, globally accessible Cloud ETL High throughput, Random I/O, SSD/Hybrid Archive High scalability, large/sequential I/O HDD Cloud Tape Hadoop / Spark Data Lakes Throughput-oriented Hybrid/HDD ML / DL Prep ⇨ Training ⇨ Inference High throughput, low latency, Random I/O SSD/NVMe Classification & Metadata Tagging High volume, index & auto-tagging zone Fast Ingest / Real-time Analytics High throughput SSD Throughput-oriented, software defined temporary landing zone capacity tier performance tier performance & capacity Tier performance & capacity Tier performance tier capacity tier Fits Traditional and New Use Cases EDGE INGEST ORGANIZE ANALYZE INSIGHTSML / DL IBM Spectrum Scale / Storage for AI / © 2020 IBM Corporation
- 43. Machine Learning is computation Understand the problem to solve Supervised vs. unsupervised learning Feature engineering - know your data Iterative methodology Greater accuracy with Deep Learning Value gained in deployments Infrastructure matters In summary
- 44. Co-Creation Lab Work side-by-side with IBM data scientists, cloud and infrastructure experts Build Custom AI Solution Plan for a pre-built Solution Design a scalable, industrial strength, training & inferencing enterprise platform IBM Systems WW Client Experience Centers: AI Center of Competency Fast Start / Design Workshops Discovery Workshop Overview Technologies Understand Business Challenges Prioritize Use Cases Showcases / Demos AI Immersion Experience AI Consulting Collateral / Assets Field Support Workshops & Co-Creation Labs DeployAdoptLearn IBM Systems ML/DL Hands-On and Customizable Combination of Lectures And Labs IBM or Customer Location IBM Portfolio and Open Source Tools Contact us at [email protected] Learn it Design it Build it Together Use it Fast Start / Design Workshops
- 45. Thank You