HTM Theory
Download as PPT, PDF6 likes9,677 views
The document provides an introduction to hierarchical temporal memory (HTM) theory and technology. It begins with a disclaimer noting that the speaker has limited knowledge and requests feedback to correct any mistakes. It then presents points related to cricket to introduce HTM concepts. The document discusses HTM algorithms, networks, properties of problems well-suited to HTM, and existing tools like NuPIC for applying HTM approaches.
HTM Theory
- 1. Md. Asfak Mahamud [email protected] Software Engineer KAZ Software Ltd. Bangladesh Date: Feb 22, 2011 Introduction to HTM
- 2. Disclaimer This theory/technology is a result from research of several decades. This theory/technology is doing a great job in various practical fields today. Many companies (for more info http://numenta.com/about-numenta/customers.php ) are using it actively. Many (e.g. http://www.atl.lmco.com/papers/1597.pdf) are doing research on its various applications. The speaker has been reading on this technology for a very short time. Any misleading or wrong explanation may occur due to lack of knowledge of the speaker. Every slide on this presentation has been tried to serve with source of reference (if required). For any confusion audiences are requested to visit the reference pages. The speaker hereby is declaring earnest eagerness to change any content of this presentation if it violets any code of conduct of any organization or company or institutions or hurts any individual.
- 3. Points • Ground. • Rule. • Ball. • Bat. • Umpire. • Player. • Powerplay. This is the first time Bangladesh is one of the organizing countries of ICC World Cup Cricket. To keep a remembrance of on going cricket craze, points are intentionally coined from sports more specifically from Cricket. Image url: http://en.wikipedia.org/wiki/File:2011_Cricket_World_Cup_Logo.svg
- 4. Points • Ground. (A rational motive for a belief or action - wordweb) • Rule. • Ball. • Bat. • Umpire. • Player. • Powerplay. Image source: http://www.cricket-for-parents.com/cricket-fielding-positions.html
- 5. At 19 Weeks of Fetus Baby's brain designates specialized areas for – smell, taste, hearing, vision, and touch. Some research suggests that baby may be able to hear sound now. Reference: http://www.babycenter.com/fetal-development-images-19-weeks Image collected and edited from: http://www.babycenter.com/fetal-development-images-19-weeks
- 6. Aniya’s Learning At her two years of old my niece, Aniya somehow knew about dog (in bengali কুকুর, ‘Kukur’). In her voice it was “Kukun”, কুকুন। She also knew the bengali word বড় ( বড়, ‘Boro’ means Big). In her voice it was ‘Bolo’, বোলো। কুকুন Image: http://jaagruti.org/2010/04/18/the-indian-street-dog/ http://www.impactlab.net/2007/04/13/building-brainlike-computers/ Image: http://www.travelblog.org/Photos/1654785 বোলো কুকুন Supervised Unsupervised
- 7. Ambiguity Vernon B Mountcastle Formerly University Professor of Neuroscience, Johns Hopkins University, Baltimore, Maryland, USA Image: http://www.ibro.org/Pub/Pub_Main_Display.asp?LC_Docs_ID=3561 An organizing principle for cerebral function: the unit module and the distributed system. In The Mindful Brain. MIT Press, 1978. Info: Video: http://tinyurl.com/4n9ux59 Dileep George, Cognitive Computing, 2007 Paper: http://tinyurl.com/464x4yt Dileep George, Phd Thesis, Stanford University, June, 2008 Same Algorithm for - Audition, - Vision, - Speech and so on. Common Cortical Algorithm No learning algorithm has an inherent superiority over another algorithm for all learning problems. (Wolpert, 1995) No Free Lunch Theorem “An algorithm’s superiority comes from the assumptions that it makes about the problem at hand.”
- 8. Solution To find Assumptions – General enough to large classes of problems. – Specific enough to make learning possible. Info from: Video: http://tinyurl.com/4n9ux59 Dileep George, Cognitive Computing, 2007 To believe 3 three things • The principles of brain function can be understood. • We can build machines that work on these principles. • Many machine learning, AI and robotics problems can only be solved this way. Jeff Hawkins, November 2010, “Advances in Modeling Neocortex and its impacts on machine intelligence” Link: http://tinyurl.com/4reyyzf World Cortex Physics Statistics Biology Anatomy Structure Physiological Results Psychological Results Info from: video: http://tinyurl.com/4n9ux59 Dileep George, Cognitive Computing, 2007
- 9. Comparison of neocortex among mouse, monkey and human. The neocortical surfaces are colored blue. Image ref: http://www.nibb.ac.jp/brish/Gallery/cortexE.html Why Neocortex is on focus? Felleman and Van Essens (1991) Model of the cortical hierarchy Image from: http://thebrain2.wikidot.com/tribal-networks • 75% of volume of human brain • All high level vision, audition, motor, language, thought. • Composed of a repetitive element – Complex – Hierarchical Jeff Hawkins, November 2010, “Advances in Modeling Neocortex and its impacts on machine intelligence” Link: http://tinyurl.com/4reyyzf
- 10. Cells that are vertically aligned in columns all respond to edges with the same orientation. Figure: Response properties of cells in V1, the first cortical region to process information from the retina. Ref: Appendix A: A Comparison between Biological Neurons and HTM cells Appendix B: A Comparison of Layers in the Neocortex and an HTM Region Hierarchical Temporal Memory including HTM Cortical Learning Algorithms http://tinyurl.com/4pr59bv Biological Neuron and Layers in Neocortex ApicalDendrite DistalDendrite ProximalDendrite Axon Layered and Columnar organization of the neocortex becomes evident when neural tissue is stained. Mini-Column: The smallest columnar Structure of the neocortex. Diameter: 30um (Approx ) Contains: 80-100 neurons across all five cellular layers.
- 11. Inter Region Major Connections Ref: Appendix B: A Comparison of Layers in the Neocortex and an HTM Region Hierarchical Temporal Memory including HTM Cortical Learning Algorithms http://tinyurl.com/4pr59bv Indirect Feed forward Pathway. Direct Feed forward Pathway. Feed back Pathway.
- 12. Biological Neuron, Simple Artificial Neuron, HTM Cell, HTM Regions Ref: Appendix A: A Comparison between Biological Neurons and HTM cells Hierarchical Temporal Memory including HTM Cortical Learning Algorithms http://tinyurl.com/4pr59bv DistalDendrite Proximal Dendrite Axon Jeff Hawkins, November 2010, “Advances in Modeling Neocortex and its impacts on machine intelligence” Link: http://tinyurl.com/4reyyzf “HOW THE BRAIN MIGHT WORK: A HIERARCHICAL AND TEMPORAL MODEL FOR LEARNING AND RECOGNITION”, Dileep George, June 2008, Stanford University
- 13. Old Wine in New Glass? NO !! Hierarchy in space and time. Slowness of time, combined with the hierarchy, enables efficient learning of intermediate levels of the hierarchy. Learning of causes by using time continuity and actions. Models of attention and specific memories. A probabilistic model specified in terms of relations between a hierarchy of causes. Belief Propagation in the hierarchy to use temporal and spatial context for inference. Ref: http://numenta.com/htm-overview/education/HTM_Comparison.pdf
- 14. Old Wine in New Glass? Ref: http://numenta.com/htm-overview/education/HTM_Comparison.pdf General- Purpose Probabilistic Models • Baysian Networks • Energy-based Models (HTMs do not treat them as rivals but tools in its toolbox.) • HHHMs (Hierarchical Hidden Markov Model, a special form of Baysian Network) have hierarchy only in space. HTM has hierarchy in space and time. • Boltzman Machine and Helhmholtz Machine do not include temporal aspects of data in the model and do not make any assumptions about hierarchy. Non- Generative Models • Support Vector Machines (SVMs) • Classic Neural Networks • Slow Feature Analysis (Many properties similar to HTM) • They are typically supervised. HTMs are fundamentally unsupervised. • They are not able to generate data for predictions. HTM can do that. Empirical Neurobiological Model • HMAX Model (Many properties similar to HTMs) • It can not predict forward in time. HTM can do that.
- 15. Old Wine in New Glass? Ref: http://numenta.com/htm-overview/education/HTM_Comparison.pdf General- Purpose Probabilistic Models • Baysian Networks • Energy-based Models (HTMs do not treat them as rivals but tools in its toolbox.) • HHHMs (Hierarchical Hidden Markov Model, a special form of Baysian Network) have hierarchy only in space. HTM has hierarchy in space and time. • Boltzman Machine and Helhmholtz Machine do not include temporal aspects of data in the model and do not make any assumptions about hierarchy. Non- Generative Models • Support Vector Machines (SVMs) • Classic Neural Networks • Slow Feature Analysis (Many properties similar to HTM) • They are typically supervised. HTMs are fundamentally unsupervised. • They are not able to generate data for predictions. HTM can do that. Empirical Neurobiological Model • HMAX Model (Many properties similar to HTMs) • It can not predict forward in time. HTM can do that.
- 16. Old Wine in New Glass? Ref: http://numenta.com/htm-overview/education/HTM_Comparison.pdf General- Purpose Probabilistic Models • Baysian Networks • Energy-based Models (HTMs do not treat them as rivals but tools in its toolbox.) • HHHMs (Hierarchical Hidden Markov Model, a special form of Baysian Network) have hierarchy only in space. HTM has hierarchy in space and time. • Boltzman Machine and Helhmholtz Machine do not include temporal aspects of data in the model and do not make any assumptions about hierarchy. Non- Generative Models • Support Vector Machines (SVMs) • Classic Neural Networks • Slow Feature Analysis (Many properties similar to HTM) • They are typically supervised. HTMs are fundamentally unsupervised. • They are not able to generate data for predictions. HTM can do that. Empirical Neurobiological Model • HMAX Model (Many properties similar to HTMs) • It can not predict forward in time. HTM can do that.
- 17. HTM in the List of AI Projects Some AI Projects Brain simulation • Blue Brain Project, HNeT (Holographic Neural Technology), Hierarchical Temporal Memory Cognitive architectures • CALO , SHIAI (Semi Human Instinctive Artificial Intelligence) , Virtual Woman Games • Chinook, Deep Blue , FreeHAL, TD-Gammon Knowledge and reasoning • Cyc , Eurisko, Open Mind Common Sense, Questsin, SNePS, Watson. Motion and manipulation • Cog, Grand Challenge 5 Natural language processing • AIML, A.L.I.C.E., ELIZA, InfoTame, Jabberwacky, KAR-Talk, PARRY, Proverb, SHRDLU, START, CSAIL, SYSTRAN, Texai Planning • O-Plan. Ref: http://en.wikipedia.org/wiki/List_of_artificial_intelligence_projects
- 18. Points • Ground. (A rational motive for a belief or action - wordweb) • Rule. (Theory/technology) • Ball. • Bat. • Umpire. • Player. • Powerplay. Image: http://blogs.trb.com/news/specials/newsillustrated/blog/2009/08/lauderhill_stadium_lets_play_c.html
- 19. HTM Algorithm HTM Cortical Learning Algorithm Jeff Hawkins, November 2010, “Advances in Modeling Neocortex and its impacts on machine intelligence” Link: http://tinyurl.com/4reyyzf Zeta 1: First generation node algorithms Image: “HOW THE BRAIN MIGHT WORK: A HIERARCHICAL AND TEMPORAL MODEL FOR LEARNING AND RECOGNITION “, Dileep George, 2008, Stanford University
- 20. HTM Network Zeta 1: First generation node algorithms Image: “HOW THE BRAIN MIGHT WORK: A HIERARCHICAL AND TEMPORAL MODEL FOR LEARNING AND RECOGNITION “, Dileep George, 2008, Stanford University
- 21. HTM Cortical Learning Algorithm Some Terminologies – Cell States • 3 output states – Active from feed-forward input – Active from lateral input – Inactive – Dendrite Segments • Proximal dendrite segment • Distal dendrite segments – Synapses • Potential Synapses • Permanence Reference: Chapter 2: HTM Cortical Learning Algorithms http://numenta.com/htm-overview/education/HTM_CorticalLearningAlgorithms.pdf
- 22. HTM Cortical Learning Algorithm Each HTM region does 3 things 1. Form a sparse distributed representation of the input. 2. Form a representation of the input in the context of the previous inputs. 3. Form a prediction based on the current input in the context of previous inputs. Reference: Chapter 2: HTM Cortical Learning Algorithms. http://numenta.com/htm-overview/education/HTM_CorticalLearningAlgorithms.pdf
- 23. HTM Cortical Learning Algorithm Context Representation of the input in the context of the previous inputs. Reference: Chapter 2: HTM Cortical Learning Algorithms. http://numenta.com/htm-overview/education/HTM_CorticalLearningAlgorithms.pdf The image has 70 columns. Each column has 4 cells. So it can save 470 Contexts by graying any combination of its columns’ cells. িকর্েকট বল । বাংলায় বল । I have eight oranges. I ate eggs.
- 24. SpatialPoolerTemporalPooler Region An HTM Region Idea Reference: Chapter 2: HTM Cortical Learning Algorithms. http://numenta.com/htm-overview/education/HTM_CorticalLearningAlgorithms.pdf
- 25. Input Sensory Data or Output from Lower Region Input
- 26. Input Sensory Data or Output from Lower Region Input&HTMRegion
- 27. Input Sensory Data or Output from Lower Region FeedForwardInput
- 28. Input Sensory Data or Output from Lower Region InhibitionRadius
- 29. Input Sensory Data or Output from Lower Region InhibitionRadius
- 30. Predictive State No Predictive States InputtoPredictiveState
- 31. Active State Active All Cells in Column InputtoPredictiveState
- 33. Points • Ground. (A rational motive for a belief or action - wordweb) • Rule. (Theory/technology.) • Ball. (Problems that fit.) • Bat. • Umpire. • Player. • Powerplay. Image source: http://www.gettyimages.com/detail/89127087/Photographers-Choice
- 34. Problem Properties Spatial Hierarchy Data generated by common sub-causes are highly correlated. In image, adjacent pixels more highly correlated than distant pixels. Ref: http://numenta.com/htm-overview/education/ProblemsThatFitHTMs.pdf Image: http://www.impactlab.net/2007/04/13/building-brainlike-computers/ Temporal Hierarchy Higher level causes vary more slowly compared to lower-level causes. In music: – Lower level individual notes changes very rapidly. – Notes are combined into musical phrase. – Phrases are combined into musical section. – Sections are combined into a symphony.
- 35. Points • Ground. (A rational motive for a belief or action – wordweb.) • Rule. (Theory/technology.) • Ball. (Problems that fit.) • Bat. (Existing tools.) • Umpire. • Player. • Powerplay. A display depicting the history of the cricket bat http://www.south-africa-tours-and-travel.com/cricket-south-africa.html
- 36. Numenta Legacy Software 2 Categories • NuPIC (Numenta Platform for Intelligent Computing) – Numenta Runtime Engine. – NuPIC Tools. – Vision Framework. • Vision Software – Vision demo applications • People tracking demo • Sample vision networks – Creating your own vision system • Vision Toolkit • Numenta Web Services Ref: http://numenta.com/legacysoftware.php
- 37. Points • Ground. (A rational motive for a belief or action – wordweb.) • Rule. (Theory/technology.) • Ball. (Problems that fit.) • Bat. (Existing tools.) • Umpire. (People behind.) • Player. • Powerplay. David Shepherd on dreaded Nelson. http://www.espncricinfo.com/magazine/content/story/149880.html
- 38. Numenta Leadership Team Subutai Ahmad VP of Engineering BS, Computer Science, Cornell University PhD, Computer Science, University of Illinois, Urbana-Champaign. Donna Dubinsky Founder, Chief Executive Officer, Board Chair B.A. Yale University, History. M.B.A., Harvard Business School. Jeff Hawkins Founder Co-founder: Palm and Handspring, Architect: PalmPilot and Treo smartphone. Book: On Intelligence. (2004) With Dileep George and Donna Dubinsky, founded Numenta in 2005. B.S. Electrical Engineering, Cornell University, 1979. Elected to the National Academy of Engineering in 2003. Dileep George Founder Led the development of the first generation of algorithms for Numenta's HTM technology. 2005 - 2010 Stanford UniversityPh.D in Electrical Engineering, 2008 Redwood Neuroscience InstituteResearch Fellow, 2003 - 2005
- 39. Numenta Board of Directors Donna Dubinsky Founder Jeff Hawkins Founder Ed Colligan Former President & Chief Executive Officer, Palm, Inc Mike Farmwald General Partner, Skymoon Ventures Harry Saal Chairman of the Technical Committee USDOJ v. Microsoft Consent Decree
- 40. Gill Bejerano Assistant Professor Developmental Biology and Computer Science Stanford University James J. DiCarlo M.D., Ph.D. Associate Professor of Neuroscience Massachussetts Institute of Technology William T. Freeman Professor Electrical Engineering and Computer Science Massachussetts Institute of Technology Andrew Y. Ng Assistant Professor Computer Science Stanford University Tomaso A. Poggio Eugene McDermott Professor, McGovern Institute Massachussetts Institute of Technology Ref: http://numenta.com/about-numenta/people.php Numenta Technical Advisory Board
- 41. Points • Ground. (A rational motive for a belief or action – wordweb.) • Rule. (Theory/technology.) • Ball. (Problems that fit.) • Bat. (Existing tools.) • Umpire. (People behind.) • Player. (Customers) • Powerplay. Bangladesh Cricket Team on a Victory Lap Image: http://www.criclounge.com/news/1637/New-Zealand-is-Bangla-washed.
- 42. Some Customers http://www.vitamindinc.com http://www.edsa.com/ http://www.forbes.com/ Other development partners are working on - detecting financial fraud, - creating new web analytical tools, and - analyzing images for digital pathology. To learn more: video from Numenta's 2009 HTM Workshop. For more info: http://numenta.com/about-numenta/customers.php
- 43. EDSA Power Analytics http://www.edsa.com/ Ref: http://www.edsa.com/pa_articles/self_learning.php
- 44. http://www.atl.lmco.com/papers/1597.pdf Lockheed Martin has been using and modifying HTM technology for several applications
- 45. http://www.atl.lmco.com/papers/1597.pdf Multimodal Pattern Recognition with Hierarchical Temporal Memory (MPR-HTM)
- 46. Points • Ground. (A rational motive for a belief or action – wordweb.) • Rule. (Theory/technology.) • Ball. (Problems that fit.) • Bat. (Existing tools.) • Umpire. (People behind.) • Player. (Customers) • Powerplay. (Demo) Image: http://www.cricketupdates.org/what-is-meant-by-powerplaypp-in-cricket.html
- 47. Application Demo vitamind Site: www.vitamindinc.com White Paper: http://www.vitamindinc.com/downloads/Vitamin%20D%20white%20paper.pdf For press review: http://www.vitamindinc.com/press.html
- 48. Application Demo vitamind Site: www.vitamindinc.com White Paper: http://www.vitamindinc.com/downloads/Vitamin%20D%20white%20paper.pdf For press review: http://www.vitamindinc.com/press.html Vitamind Outlook Picasa Webcam/Network Cam People Detected Mail, With detection image By macro attached images from the mails are saved in a directory.
- 49. Tools Demo • NuPIC (Numenta Platform for Intelligent Computing) • Numenta Vission Software Vision Framework demo Vision Toolkit demo
- 50. Java Demo Courtesy: teddybot http://numenta.com/phpBB2/viewtopic.php?t=1368 vincentvanrijn Courtesy: vincentvanrijn http://numenta.com/phpBB2/viewtopic.php?t=1368
- 51. Hierarchical Temporal Memory • Ground. (A rational motive for a belief or action – wordweb.) • Rule. (Theory/technology.) • Ball. (Problems that fit.) • Bat. (Existing tools.) • Umpire. (People behind.) • Player. (Customers) • Powerplay. (Demo) Let’sPlay