Intelligent Information Technology Research Lab, Acadia University, Canada1Daniel L. SilverAcadia University,Wolfville, NS...
Intelligent Information Technology Research Lab, Acadia University, Canada2
Intelligent Information Technology Research Lab, Acadia University, CanadaKey Take Away A major challenge in artificial i...
Intelligent Information Technology Research Lab, Acadia University, CanadaOutline Learning – What is it? History of Mach...
Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning? Animals and Humansa. Learn us...
Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning?(A little more formally) Induc...
Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning? Generalization through learni...
Intelligent Information Technology Research Lab, Acadia University, Canada9Inductive BiasASH STTHI RDSEC ONDELM STFIR STPI...
Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning? Requires an inductive bias= a...
Intelligent Information Technology Research Lab, Acadia University, CanadaInductive Biases Universal heuristics - Occam’s...
Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Machine Learning?The study of how to bu...
Intelligent Information Technology Research Lab, Acadia University, CanadaHistory of Machine Learning1950 20001980PDP Grou...
Intelligent Information Technology Research Lab, Acadia University, CanadaOf Interest to Several Disciplines Computer Sci...
Intelligent Information Technology Research Lab, Acadia University, CanadaClasses of ML Methods Supervised – Develops mod...
Intelligent Information Technology Research Lab, Acadia University, CanadaFocus: Supervised Learning Function approximati...
Intelligent Information Technology Research Lab, Acadia University, Canada23Supervised Machine LearningFrameworkInstance S...
Intelligent Information Technology Research Lab, Acadia University, CanadaSupervised Machine Learning Problem: We wish to...
Intelligent Information Technology Research Lab, Acadia University, CanadaClassificationMistakesTyping SpeedABBBBBBBBBBBBB...
Intelligent Information Technology Research Lab, Acadia University, CanadaClassificationABBBBBBBBBBBBBBBBB BBBAAAAAAAAAAAA...
Intelligent Information Technology Research Lab, Acadia University, CanadaClassificationABBBBBBBBBBBBBBBBB BBBAAAAAAAAAAAA...
Intelligent Information Technology Research Lab, Acadia University, CanadaApplicationAreasData Mining: Science and medici...
Intelligent Information Technology Research Lab, Acadia University, CanadaApplication Areas Web mining – information filt...
Intelligent Information Technology Research Lab, Acadia University, CanadaRecent and Future Advances Robotics Neuroprost...
Intelligent Information Technology Research Lab, Acadia University, CanadaOASIS: Onboard AutonomousScience Investigation S...
Intelligent Information Technology Research Lab, Acadia University, Canada Stanford’s Sebastian Thrun holds a $2M check o...
Intelligent Information Technology Research Lab, Acadia University, CanadaThe Competition36
Intelligent Information Technology Research Lab, Acadia University, CanadaAutonomous Underwater VehiclesArctic ExplorerAU...
Intelligent Information Technology Research Lab, Acadia University, CanadaLiterally Extending Our Reach– Neuroprosthetic D...
Intelligent Information Technology Research Lab, Acadia University, Canada40Lifelong Machine Learning (LML) Considers met...
Intelligent Information Technology Research Lab, Acadia University, Canada41Supervised Machine LearningFrameworkInstance S...
Intelligent Information Technology Research Lab, Acadia University, Canada42Lifelong Machine LearningFrameworkInstance Spa...
Intelligent Information Technology Research Lab, Acadia University, Canada43Lifelong Machine LearningFrameworkInstance Spa...
Intelligent Information Technology Research Lab, Acadia University, Canada44Lifelong Machine LearningOne ImplementationIns...
Intelligent Information Technology Research Lab, Acadia University, Canada48An Environmental ExampleStream flow rate predi...
Intelligent Information Technology Research Lab, Acadia University, CanadaLifelong Machine Learningwith csMTLExample: Lea...
Intelligent Information Technology Research Lab, Acadia University, CanadaLifelong Machine Learningwith csMTL55Demo
Intelligent Information Technology Research Lab, Acadia University, CanadaDeep Learning Architectures Hinton and Bengio (...
Intelligent Information Technology Research Lab, Acadia University, CanadaDeep Learning Architectures Consider the proble...
Intelligent Information Technology Research Lab, Acadia University, CanadaDeep Learning Architectures2000 top-level artifi...
Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing Power Moores Law Expected toa...
Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing Power IBMs Watson – Jeopardy, ...
Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing PowerAndrew Ng’s work on DeepLe...
Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing PowerResults: A face detector ...
Intelligent Information Technology Research Lab, Acadia University, CanadaNever-Ending Language Learner Carlson et al (20...
Intelligent Information Technology Research Lab, Acadia University, CanadaCloud-Based ML - Google69https://developers.goog...
Intelligent Information Technology Research Lab, Acadia University, CanadaMachine Flight vs.Machine Learning71Factor Machi...
Intelligent Information Technology Research Lab, Acadia University, Canada72Thank You! danny.silver@acadiau.ca http://pl...
Upcoming SlideShare
Loading in...5
×

Getting a Machine to Learn: Extending Our Reach Beyond Our Grasp

360

Published on

Machine Learning is the study of how to build systems that can automatically learn and improve with experience similar to humans. Since the early 1980’s there have been significant advances in ML that have affected things such as marketing, banking, manufacturing, household appliances, automobiles, medicine and health care, and most recently the Internet and mobile devices. Machine Learning is poised to extend human mental reach in the virtual world of the 21st century in the same way as flight extended our physical reach in 20th century – it provides the means to filter massive amounts of data, recognize complex patterns, and rapidly make difficult decisions.


Danny is a Professor in and the Director of the Jodrey School of Computer Science at Acadia University. His areas of research and development are machine learning, data mining, and adaptive systems. He has published over 60 scientific papers, edited special journal editions, and has been part of the program committee for a number of national and international conferences, seminars and workshops. Most recently he was awarded a Harrison McCain Foundation Award for research into advance machine learningmethods. Since 1993, he has worked on machine learning and data mining projects in the private and public sector providing situation analysis and problem definition, project management and guidance, and predictive analytic services. In 2011, he received the Science Champion Award from the Nova Scotia Discovery Center for his work on youth robotics and the advancement of STEM education.

Published in: Technology, Education
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
360
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
0
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide
  • Show OH of multi-disciplinary nature of study of ANNs
  • Mention RASL3 here, show the methods of KT again, name outputs T4-T8
  • Transcript of "Getting a Machine to Learn: Extending Our Reach Beyond Our Grasp"

    1. 1. Intelligent Information Technology Research Lab, Acadia University, Canada1Daniel L. SilverAcadia University,Wolfville, NS, Canadadanny.silver@acadiau.ca
    2. 2. Intelligent Information Technology Research Lab, Acadia University, Canada2
    3. 3. Intelligent Information Technology Research Lab, Acadia University, CanadaKey Take Away A major challenge in artificial intelligence hasbeen how to develop common backgroundknowledge Machine learning systems are beginning tomake head-way in this area Taking first steps to captureknowledge that can be usedfor future learning, reasoning,etc.3
    4. 4. Intelligent Information Technology Research Lab, Acadia University, CanadaOutline Learning – What is it? History of Machine Learning Framework and Methods ML Application Areas Recent and Future Advances Challenges and Open Questions4
    5. 5. Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning? Animals and Humansa. Learn using new experiences and priorknowledgeb. Retain new knowledge from what is learnedc. Repeat starting at 1. Essential to our survival and thriving5
    6. 6. Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning?(A little more formally) Inductive inference/modeling Developing a general model/hypothesis fromexamples Objective is to achieve good generalization formaking estimates/predictions It’s like … Fitting a curve to data Also considered modeling the data Statistical modeling7
    7. 7. Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning? Generalization through learning is notpossible without an inductive bias= a heuristic beyond the data
    8. 8. Intelligent Information Technology Research Lab, Acadia University, Canada9Inductive BiasASH STTHI RDSEC ONDELM STFIR STPINE STOAK STInductive bias depends upon:• Having prior knowledge• Selection of most relatedknowledgeHuman learners use Inductive Bias
    9. 9. Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Learning? Requires an inductive bias= a heuristic beyond the data Do you know any inductive biases? How do you choose which to use?
    10. 10. Intelligent Information Technology Research Lab, Acadia University, CanadaInductive Biases Universal heuristics - Occam’s Razor Knowledge of intended use – Medicaldiagnosis Knowledge of the source - Teacher Knowledge of the task domain Analogy with previously learned tasksTom Mitchell, 1980
    11. 11. Intelligent Information Technology Research Lab, Acadia University, CanadaWhat is Machine Learning?The study of how to build computerprograms that: Improve with experience Generalize from examples Self-program, to some extent
    12. 12. Intelligent Information Technology Research Lab, Acadia University, CanadaHistory of Machine Learning1950 20001980PDP GroupMulti-layerPerceptrons,New appsRenaissance1990AI SuccessData mining,Web mining,User models,New alg.,GooglePresentBig Data,Web Analytics,Parallel alg.,Cloud comp.,Deep learningAdvances1890WilliamJames,NeuronallearningOrigins1940Donald Hebb,Math models,The PerceptronLimited valuePromise1960Minsky &Papertpaper,ResearchwanesHiatus1970Genetic alg,Versionspaces,DecisionTreesExploration
    13. 13. Intelligent Information Technology Research Lab, Acadia University, CanadaOf Interest to Several Disciplines Computer Science – theory of computation, newalgorithms Math - advances in statistics, information theory Psychology – as models for human learning, knowledgeacquisition and retention Biology – how does a nervous system learn Physics – analogy to physical systems Philosophy – epistemology, knowledge acquisition Application Domains – new knowledge extracted fromdata, solutions to unsolved problems17
    14. 14. Intelligent Information Technology Research Lab, Acadia University, CanadaClasses of ML Methods Supervised – Develops models that predict the value ofone variable from one or more others: Artifical Neural Networks, Inductive Decision Trees, GeneticAlgorithms, k-Nearest Neighbour, Bayesian Networks, SupportVectors Machines Unsupervised – Generates groups or clusters of datathat share similar features K-Means, Self-organizing Feature Maps Reinforcement Learning – Develops models from theresults of a final outcome; eg. win/loss of game TD-learning, Q-learning (related to Markov Decision Processes) Hybrids – eg. semi-supervised learning
    15. 15. Intelligent Information Technology Research Lab, Acadia University, CanadaFocus: Supervised Learning Function approximation(curve fitting) Classification (concept learning, patternrecognition)x1x2ABf(x)x21
    16. 16. Intelligent Information Technology Research Lab, Acadia University, Canada23Supervised Machine LearningFrameworkInstance SpaceXTrainingExamplesTestingExamples(x, f(x))Model ofClassifierhInductiveLearning Systemh(x) ~ f(x)
    17. 17. Intelligent Information Technology Research Lab, Acadia University, CanadaSupervised Machine Learning Problem: We wish to learn to classifying two people(A and B) based on their keyboard typing. Approach: Acquire lots of typing examples from each person Extract relevant features - representation! M = number of mistakes T = typing time Transform feature representation as needed Use an algorithm to fit a model to the data - search! Test the model on an independent set of examples of typing fromeach person
    18. 18. Intelligent Information Technology Research Lab, Acadia University, CanadaClassificationMistakesTyping SpeedABBBBBBBBBBBBBBBBB BBBAAAAAAAAAAAAAAAAAAABBBBBBBBBLogistic RegressionYY=f(M,T)01M TY
    19. 19. Intelligent Information Technology Research Lab, Acadia University, CanadaClassificationABBBBBBBBBBBBBBBBB BBBAAAAAAAAAAAAAAAAAAABBBBBBBBBArtificial Neural NetworkAMistakesTyping SpeedM TY…
    20. 20. Intelligent Information Technology Research Lab, Acadia University, CanadaClassificationABBBBBBBBBBBBBBBBB BBBAAAAAAAAAAAAAAAAAABBBBBBBBBInductive Decision TreeAAMistakesTyping SpeedM?T? T?RootLeafABBlood Pressure Example
    21. 21. Intelligent Information Technology Research Lab, Acadia University, CanadaApplicationAreasData Mining: Science and medicine: prediction, diagnosis, patternrecognition, forecasting Manufacturing: process modeling and analysis Marketing and Sales: targeted marketing, segmentation Finance: portfolio trading, investment support Banking & Insurance: credit and policy approval Security: bomb, iceberg, fraud detection Engineering: dynamic load shedding, pattern recognition31
    22. 22. Intelligent Information Technology Research Lab, Acadia University, CanadaApplication Areas Web mining – information filtering and classification,social media predictive modeling User Modeling – adaptive user interfaces,speech/gesture recognition Intelligent Personal Agents – email spamfiltering, fashion consultant, Robotics – image recognition, adaptive control,autonomous vehicles (space, under-sea) Military/Defense – target acquisition and classification,tactical recommendations, cyber attack detection32
    23. 23. Intelligent Information Technology Research Lab, Acadia University, CanadaRecent and Future Advances Robotics Neuroprosthetics Lifelong Machine Learning Deep Learning Architectures ML and Growing Computing Power NELL – Never-Ending Language Learner Cloud-based Machine Learning33
    24. 24. Intelligent Information Technology Research Lab, Acadia University, CanadaOASIS: Onboard AutonomousScience Investigation System Since early 2000’s Goal: To evaluate,and autonomouslyact upon, sciencedata gathered byspacecraft Including planetarylanders and rovers34
    25. 25. Intelligent Information Technology Research Lab, Acadia University, Canada Stanford’s Sebastian Thrun holds a $2M check on top ofStanley, a robotic Volkswagen Touareg R5 212 km autonomus vehicle race, Nevada Stanley completed in 6h 54m Four other teams also finishedSource: Associated Press – Saturday, Oct 8, 2005DARPA GrandChallenge 200535
    26. 26. Intelligent Information Technology Research Lab, Acadia University, CanadaThe Competition36
    27. 27. Intelligent Information Technology Research Lab, Acadia University, CanadaAutonomous Underwater VehiclesArctic ExplorerAUV designed and built by InternationalSubmarine Engineering Ltd. (ISE) of PortCoquitlam, B.C.Used to map the sea floor underneath theArctic ice shelf in support of Canadian landclaims under the UN Convention on theLaw of the Sea.Various military uses; e.g. mine detection,elimination(Source: ISE, Mae Seto)37
    28. 28. Intelligent Information Technology Research Lab, Acadia University, CanadaLiterally Extending Our Reach– Neuroprosthetic Decoders Dec, 2012 Andy Schwart,Univ. of Pittsburgh Jan Scheuermann,quadriplegic Brain-machineinterface, 96electrodes 13 weeks oftraining High-performance neuroprostheticcontrol by an individual with tetraplegia,The Lancet, v381, p557-654, Feb 201339
    29. 29. Intelligent Information Technology Research Lab, Acadia University, Canada40Lifelong Machine Learning (LML) Considers methods of retaining and usinglearned knowledge to improve the effectivenessand efficiency of future learning We investigate systems that must learn: From impoverished training sets For diverse domains of tasks Where practice of the same task happens Applications: Intelligent Agents, Robotics, User Modeling, DM
    30. 30. Intelligent Information Technology Research Lab, Acadia University, Canada41Supervised Machine LearningFrameworkInstance SpaceXTrainingExamplesTestingExamples(x, f(x))Model ofClassifierhInductiveLearning Systemh(x) ~ f(x)After model is developedand used it is thrown away.
    31. 31. Intelligent Information Technology Research Lab, Acadia University, Canada42Lifelong Machine LearningFrameworkInstance SpaceXTrainingExamplesTestingExamples(x, f(x))Model ofClassifierhInductiveLearning Systemshort-term memoryh(x) ~ f(x)DomainKnowledgelong-term memoryRetention &ConsolidationInductiveBias SelectionKnowledgeTransfer
    32. 32. Intelligent Information Technology Research Lab, Acadia University, Canada43Lifelong Machine LearningFrameworkInstance SpaceXTrainingExamplesTestingExamples(x, f(x))Model ofClassifierhInductiveLearning Systemshort-term memoryh(x) ~ f(x)DomainKnowledgelong-term memoryRetention &ConsolidationInductiveBias SelectionKnowledgeTransfer
    33. 33. Intelligent Information Technology Research Lab, Acadia University, Canada44Lifelong Machine LearningOne ImplementationInstance SpaceXTrainingExamplesTestingExamples(x, f(x))Model ofClassifierhh(x) ~ f(x)Retention &ConsolidationKnowledgeTransferf2(x)x1 xnf1(x) f5(x)Multiple TaskLearning (MTL)InductiveBias Selectionf3(x)f2(x) … f9(x) fk(x)ConsolidatedMTLDomainKnowledgelong-term memory
    34. 34. Intelligent Information Technology Research Lab, Acadia University, Canada48An Environmental ExampleStream flow rate prediction [Lisa Gaudette, 2006]x = weather dataf(x) = flow rate1112131415160 1 2 3 4 5 6Years of Data TransferedMAE(m^3/s)No Transfer Wilmot Sharpe Sharpe & Wilmot Shubenacadie
    35. 35. Intelligent Information Technology Research Lab, Acadia University, CanadaLifelong Machine Learningwith csMTLExample: Learning to Learn howto transform images Requires methods ofefficiently & effectively Retaining transformmodel knowledge Using this knowledge tolearn new transforms(Silver and Tu, 2010)52
    36. 36. Intelligent Information Technology Research Lab, Acadia University, CanadaLifelong Machine Learningwith csMTL55Demo
    37. 37. Intelligent Information Technology Research Lab, Acadia University, CanadaDeep Learning Architectures Hinton and Bengio (2007+) Learning deep architectures of neuralnetworks Layered networks of unsupervised auto-encoders efficiently develop hierarchiesof features that capture regularities intheir respective inputs Used to develop models for families of tasks57
    38. 38. Intelligent Information Technology Research Lab, Acadia University, CanadaDeep Learning Architectures Consider the problem of trying to classifythese hand-written digits.
    39. 39. Intelligent Information Technology Research Lab, Acadia University, CanadaDeep Learning Architectures2000 top-level artificial neurons2000 top-level artificial neurons00500 neurons(higher level features)500 neurons(higher level features)500 neurons(low level features)500 neurons(low level features)Images ofdigits 0-9(28 x 28 pixels)Images ofdigits 0-9(28 x 28 pixels)11 22 33 4455 66 77 88 99Neural Network:- Trained on 40,000 examples- Learns:* labels / recognize images* generate images from labels- Probabilistic in nature- Demo231
    40. 40. Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing Power Moores Law Expected toaccelerate as thepower of computersmove to a log scalewith use of multipleprocessing cores60
    41. 41. Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing Power IBMs Watson – Jeopardy, Feb, 2011: Massively parallel data processing system capableof competing with humans in real-time question-answer problems 90 IBM Power-7 servers Each with four 8-core processors 15 TB (220M text pages) of RAM Tasks divided into thousands of stand-alonejobs distributed among 80 teraflops (1 trillion ops/sec) Uses a variety of AI approaches includingmachine learning61
    42. 42. Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing PowerAndrew Ng’s work on DeepLearning Networks (ICML-2012)Problem: Learn to recognize humanfaces, cats, etc from unlabeled dataDataset of 10 million images; eachimage has 200x200 pixels9-layered locally connected neuralnetwork (1B connections)Parallel algorithm; 1,000 machines(16,000 cores) for three days62Building High-level Features Using Large Scale Unsupervised LearningQuoc V. Le, Marc’Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen,Greg S. Corrado, Jeffrey Dean, and Andrew Y. NgICML 2012: 29th International Conference on Machine Learning, Edinburgh,Scotland, June, 2012.
    43. 43. Intelligent Information Technology Research Lab, Acadia University, CanadaML and Computing PowerResults: A face detector that is 81.7%accurate Robust to translation, scaling,and rotationFurther results: 15.8% accuracy in recognizing20,000 object categories fromImageNet 70% relative improvement overthe previous state-of-the-art.63
    44. 44. Intelligent Information Technology Research Lab, Acadia University, CanadaNever-Ending Language Learner Carlson et al (2010) Each day: Extracts information from theweb to populate a growing knowledgebase of language semantics Learns to perform this task better than onprevious day Uses a MTL approachin which a large numberof different semanticfunctions are trainedtogether64
    45. 45. Intelligent Information Technology Research Lab, Acadia University, CanadaCloud-Based ML - Google69https://developers.google.com/prediction/
    46. 46. Intelligent Information Technology Research Lab, Acadia University, CanadaMachine Flight vs.Machine Learning71Factor Machine Flight Machine LearningEffectiveness Travel higher, father Learn more things, accuratelyTo places not reachable Model complex phenomenaEfficiency Travel faster Learn fasterLower cost Lower costSatisfaction Safe travel, beauty Confidence, eleganceReach the moon,and beyondReach new knowledge,solve new problems
    47. 47. Intelligent Information Technology Research Lab, Acadia University, Canada72Thank You! danny.silver@acadiau.ca http://plato.acadiau.ca/courses/comp/dsilver/ http://ML3.acadiau.ca

    ×