What's The Role Of Machine Learning In Fast Data And Streaming Applications?
6 likes1,965 views
The document discusses the role of machine learning (ML) in enhancing fast data and streaming applications, highlighting its ability to improve efficiency in tasks like video analysis and medical diagnosis. It outlines various use cases of ML in transformative technologies and emerging trends, emphasizing the democratization of ML tools for wider accessibility. Additionally, it presents the machine learning lifecycle and relevant tools for development and operational management, ultimately demonstrating how ML integration can drive innovation in diverse sectors.
What's The Role Of Machine Learning In Fast Data And Streaming Applications?
- 1. What's The Role of Machine Learning In Fast Data and Streaming Applications? WEBINAR Emre Velipasaoglu, Ph.D Principal Data Scientist
- 2. What is Machine Learning? A computer program is a set of explicit instructions that produce an output for a given input. Machine Learning (ML) is about how to program computers to improve automatically with experience rather than explicit instructions. val x = 0 to 1000000000000 val y = x.map(x => x/2.0 + 1) 1 -> 1.5 2 -> 2.0 => 10^12 -> ?
- 3. Why does everyone want to use it? A lot of the recent transformative technologies are based on ML: • optical character recognition • speech recognition • fraud detection • web search • personalized marketing and advertising • Computer Aided medical Diagnosis (CADx)
- 4. Why does everyone want to use it? Emerging trends that will leverage ML: • Internet of Things (IoT) • Augmented Reality (AR) and Virtual Reality (VR) • Autonomous vehicles • Customer service chat bots • Security: Face/voice/biometrics recognition • Healthcare: drug discovery, outcome prediction, personalized care • Democratization of ML and the long tail of ML applications
- 5. What are some of the use cases? Machines do certain tasks better than humans. • IBM DeepBlue in chess • IBM Watson in Jeopardy • Google AlphaGo in go • Lip reading: LipNet 93% vs. humans 65%
- 6. What are some of the use cases? Machines are more cost efficient in certain tasks: • Transcribing: Microsoft 89% vs. humans 89% • Computer-aided diagnosis: E.g. Dermatologist-level classification of skin cancer with deep neural networks, Esteva, et. al., published in Nature, June 2017.
- 7. What are some of the use cases? Machines are the only way to scale up processing in certain tasks. • Previewing video: Clarifai can analyze 3.5 minutes of video in 10 seconds for detecting objects. • Commercial loan agreements review: AI in seconds vs. humans in 360,000 hours.
- 8. Why should you care? Information gives competitive advantage. ML unlocks information from your data It is your product, your data, your operations. ML is being democratized. It is not for a handful of giant tech companies anymore.
- 9. How does it work? E.g. Augmented Reality (AR) Shopping Personalization: • Shopping is one area where AR is expected to impact. • IBM CeBIT 2013 app: • scans a shelf • recognizes products • overlays nutritional info • Add a recommender system, tailored to • your customers, • your product catalog.
- 10. Learning a Recommender User Rating Matrix A B C D Alan 5 1 1 Emre 4 2 3 Vishal 5 1 Matrix Factorization User Latent Factor Model f1 f2 Alan 1.63 0.89 Emre 0.89 2.10 Vishal 2.03 1.01 Item Latent Factor Model A B C D f1 2.21 1.88 -0.24 0.33 f2 1.68 1.22 1.51 0.74
- 11. Scoring Items Query User User Latent Factors f1 f2 Vishal 2.03 1.01 Query Item Estimate Ratings Ranked Items score A 6.00 D 1.41 C 1.05 Item Latent Factors A C D f1 2.21 -0.24 0.33 f2 1.68 1.51 0.74 User Latent Factor Model f1 f2 Alan 1.63 0.89 Emre 0.89 2.10 Vishal 2.03 1.01 Item Latent Factor Model A B C D f1 2.21 1.88 -0.24 0.33 f2 1.68 1.22 1.51 0.74 ApplicationUserItems
- 12. Modern ML • Size of data • Does not fit in one node, must distribute • E.g. billions of user x item ratings (several orders of magnitude more historical events to aggregate ratings from) • Size of model • Does not fit in one node, must distribute • E.g. millions of users, thousands of products • Learning speed • Anywhere from real-time model updates to batch updates in minutes • Operational latency and throughput • Low milliseconds response time for millions of transactions
- 13. ML Lifecycle - Development Early research • Explore of modeling techniques Iterations • Feature selection • Training parameter tuning Productization • Feature computations, model updating, scoring, caching, optimizations, etc. (e.g. update and query of latent factors) recommender system collaborative filtering content-based filtering hybrid models Bayesian networks clustering latent semantic models Markov decision process … singular value decomposition alternating least squares non-negative matrix factorization number of latent factors step size convergence criteria
- 14. ML Lifecycle - Management Monitoring • Model performance • Latency • Throughput • Model quality • Drift (e.g. has the user’s tastes changed recently?) • Security and robustness Controlling • Model optimization (for performance) • Model update (for quality)
- 15. Which tools are available and what do they do? Machine Learning Spark MLlib ML library for Spark Flink ML ML library for Flink Mahout Distributed or scalable ML algorithms Tensorflow Google's open source deep learning library Theano Numerical library for Python, especially for deep learning Deeplearning4j Deep learning library in Java BigDL Intel’s distributed deep learning library on Spark scikit-learn Main ML library for Python OpenNLP ML based toolkit for the processing of natural language text
- 16. Which tools are available and what do they do? Streaming Flink Stream processing framework with sophisticated handling of late arriving data Spark Streaming Dataset based computing framework with mini-batch streaming support Beam API for data processing pipelines Data Ingestion Kafka Distributed stream processing for high-throughput, low- latency, real-time data feeds Flume Log processing
- 17. Which tools are available and what do they do? Persistence and Storage HDFS Hadoop based Distributed File System Cassandra Distributed NoSQL database management system ElasticSearch Distributed, RESTful search and analytics engine AWS S3 Cloud based object store
- 18. How does the Fast Data Platform tie it all together? HDFS User Rating Matrix Spark Matrix Factorizat ion Cassandra Latent Item Model Latent User Model Flink / Akka Streams Query Item Query User Item Factors User Factors Score Kafka Ranked Items Kafka User Items Application batch streaming
- 19. What else does FDP provide? data persistence & storage stream processing machine learning cluster analysis infrastructure durable messaging backplane microservices intelligent management
- 20. In Summary • Machine Learning is the way to build transformative products leveraging data that are otherwise impossible to build. • It is not difficult to build Machine Learning based solutions, thanks to new open source tools. • Lightbend’s Fast Data Platform provides an easy onramp for building, deploying and running Fast Data clusters and services leveraging best of breed tools .
- 21. Upgrade your grey matter! Get the free O’Reilly book by Dr. Dean Wampler, VP of Fast Data Engineering at Lightbend bit.ly/lightbend-fast-data