Introduction to Deep Learning and AI at Scale for Managers
3 likes463 views
The document provides an overview of deep learning and AI tailored for managers, highlighting key concepts such as machine learning applications, the evolution of algorithms, and the significance of data. It discusses the pros and cons of deep learning, necessary algorithms, and frameworks for implementation, as well as challenges related to debugging, bias, and accuracy. The author emphasizes the integration of AI with conventional systems and offers resources for starting AI projects.
![Approaches to Machine Learning
8
Source: S. Raschka: An Introduc@on to Supervised
Machine Learning and Pa8ern Classifica@on: The Big Picture
95% of industry use
cases
[Gartner]](https://image.slidesharecdn.com/vladimirbacvanskiintroductiontodeeplearningandaiatscaleformanagers-170621195808/85/Introduction-to-Deep-Learning-and-AI-at-Scale-for-Managers-8-320.jpg)
![§ Quora experience: [Quora]
• Spark implementa@on: 6hr, 15 machines
• C++ implementa@on: 10 min, 1 machine
§ Some others found 7x improvement 13 node distributed vs. single
machine [Databricks]
§ You may not need distributed solu@ons for every problem!
• Maybe you can reduce your data set?
• Smart sampling may be be8er
• Use if your data is really big and workload parallelizable
• Always benchmark with your data set
Distribu)ng Machine Learning: Do You Need It?
www.scispike.com 18](https://image.slidesharecdn.com/vladimirbacvanskiintroductiontodeeplearningandaiatscaleformanagers-170621195808/85/Introduction-to-Deep-Learning-and-AI-at-Scale-for-Managers-18-320.jpg)
Introduction to Deep Learning and AI at Scale for Managers
- 2. § Founder of SciSpike, a development, consul@ng, and training firm § Passionate about So5ware and Big Data § PhD in computer science RWTH Aachen, Germany § Architect, consultant, mentor Dr. Vladimir Bacvanski § Custom development and consul@ng § Scalable Web and IoT systems § Healthcare, Smart Ci@es, E-Commerce,… www.scispike.com
- 3. Ar@ficial Intelligence AI, Machine Learning, Deep Learning www.scispike.com 3 Machine Learning 1950's 1980's 2010's Deep Learning "Dog"
- 6. Star)ng With Machine Learning Popular applica)ons: § Text and speech recogni@on § Recommenda@on systems § Image classifica@on § Spam filtering Machine Learning is a study of computer algorithms that improve automa7cally through experience Tom Mitchell § Face recogni@on § Computer security § Fraud detec@on § Predic@ng future events and values
- 10. § A simple algorithm with more data may perform be8er than a complex model • Op@mize only what ma8ers § The UI is the main communica@on channel that ma8ers to users • Data à ML à UI (Visualiza@on / Speech) BeLer Models or More Data? www.scispike.com 10 "Google does not have be<er algorithms, only more data" Peter Norvig, Director of research, Google
- 11. § Deep Learning: a branch of ML that uses mul@ple (deep) processing layers, composed of mul@ple non-linear transforma@ons. • 70's Neural Networks used with much more compu@ng power • More data + bigger models + more computa@on • GPUs make it much faster Deep Learning - Deep Neural Networks www.scispike.com 11
- 12. § A single layer neural network: adjust hyperparameters un@l you get the desired results Layers in a Neural Network www.scispike.com 12 Source: h8p://deeplearning4j.org/ Number of Hidden Layers - Deep: 3 - Very deep: 16+ - Extremely deep: 50 – 1000s Most of the problems use1-2 hidden layers
- 13. Build a model from a set of examples Start with a random set of parameters Measure against the known correct output Modify parameters, try to improve the match • This is the tricky part! The model eventually makes useful predic@ons But How Does it Work? www.scispike.com 13 Deep Learning Systems: - Tensorflow - Keras - MXNet - Caffe - CNTK - Theano - …
- 14. Layers: Deepen Intermediate Representa)ons www.scispike.com 14 Source: h8p://deeplearning4j.org/
- 16. Deep Learning Pros and Cons Pros • Conceptually simple • Non-linear • Highly flexible • Can be fine-tuned with more data • Excellent for pa8ern recogni@on Cons • Hard to interpret • Theory not well understood • Slow to train • May overfit • Data hungry www.scispike.com 16
- 18. § Quora experience: [Quora] • Spark implementa@on: 6hr, 15 machines • C++ implementa@on: 10 min, 1 machine § Some others found 7x improvement 13 node distributed vs. single machine [Databricks] § You may not need distributed solu@ons for every problem! • Maybe you can reduce your data set? • Smart sampling may be be8er • Use if your data is really big and workload parallelizable • Always benchmark with your data set Distribu)ng Machine Learning: Do You Need It? www.scispike.com 18
- 20. § AI close to the source of events – Determine if the event should be sent to back-end – Only highly relevant events are processed • Pa8ern matching, classifica@on, recogni@on – Preserves the bandwidth, saves processing resources § AI on the back-end – Event data processed with AI – AI used for things to difficult for conven@onal processing – Pa8ern matching, classifica@on, recogni@on Event Sources and AI Applica)ons www.scispike.com 20
- 21. From Event Flows, Streaming to AI Enhanced Behavior www.scispike.com 21 Easy start: cloud based ML and AI services ML/AI Components: • Object and scene detec@on • Facial search and recogni@on • Language understanding • Recommenda@ons • Text analy@cs • …
- 22. § AI systems are not 100% accurate § What is the level of trust? – Accuracy 50% • We don't trust it, but may accept result as a help • We are always ready to intervene – Accuracy 99% • We learn to trust it, and don't pay a8en@on to it any more • In case of failure, humans are unlikely to respond in @me § What is the consequence of failure? Challenges of AI and Streaming: Accuracy www.scispike.com 22
- 25. § Deep Learning problems: – Confidently classifying random data – Minuscule perturba@ons of inputs may lead to misclassifica@on … and Failures www.scispike.com 25
- 26. § Uncertainty in development § Difficult to forecast success – Choice of an algorithm – Choice of the data set § The solu@ons may incorporate bias based on the data sets provided for training § Is the tes@ng data set representa@ve of real world? § Integra@on is essen@al: large parts of AI applica@ons are conven@onal Big Data / Streaming / Data management systems Challenges of Deep Learning and AI Development www.scispike.com 26
- 27. § AI (and other advanced ML models) are (mostly) black boxes § We have the understanding of mechanism deployed, but not the relevance of factors created during training § In face of errors, where do you debug? – Start training from the beginning? – Somewhere in the middle? • Focus on the last layer of deep learning systems? Challenges of Deep Learning and AI Debugging www.scispike.com 27
- 28. § Look for inspira@on in publicly available machine learning applica@ons § Tools and libraries reduce (but do not eliminate!) the need for specialized talent Business Innova)on with DL and AI www.scispike.com 28 Data Research and Build Hypotheses Build ML Solu7ons AB tes7ng, Real world confirma7on Data Science Data Science ML Engineering
- 29. • Iden@fy the areas where Deep Learning and AI can add value • What is the value for the customer? • Dis@nguish short term experiment vs. long term goal • Integrate Deep Learning and AI with Big Data systems • Not ge7ng started: a mistake • Get your hands dirty! • If you are a business person: pair up with a curious developer! Deep Learning And AI in Your Organiza)on www.scispike.com 29
- 30. § Start with "A visual introduc@on to machine learning" – h8p://www.r2d3.us/visual-intro-to-machine-learning-part-1/ § Great insight into AI: Columbia University Ar@ficial Intelligence course – h8ps://courses.edx.org/courses/course-v1:ColumbiaX+CSMM.101x+2T2017/ – For managers: watch the videos, you can skip the labs § Amazon, Google, Micorso5 AI services – variety of services § Some other services: – Wit.ai: turn voice into text, execute commands – api.ai: conversa@onal user experience: build bots – clarifai.com: visual recogni@on § Tensorflow – a popular deep learning framework. Keras makes it easier – h8ps://www.tensorflow.org/, h8ps://keras.io/ How To Start Your First AI Project: Resources www.scispike.com 30