Lifted Message Passing Rorschach Test K. Kersting 2 Lifted - PDF document

GRAPHBOT 2010. IROS � WORKSHOP ON PROBABILISTIC GRAPHICAL Sriraam Babak Fabian Scott Youssef El Natarajan Ahmadi Hadiji Sanner Massaoudi MODELS IN ROBOTICS Taipei, Taiwan, October 22. 2010 A E Kristian Kersting Lifted Message Passing Rorschach Test K. Kersting 2 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 1

Etzioni’s Rorschach Test for Computer Scientists 3 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Moore’s Law? K. Kersting 4 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 2

Storage Capacity? 5 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Number of Facebook Users? K. Kersting 6 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 3

Number of Scientific Publications? 7 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Number of Web Pages? K. Kersting 8 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 4

Number of Actions? 9 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Computing 2020: Science in an Exponential World “The amount of scientific data is doubling every year” [Szalay,Gray; ������ 440, 413�414 (23 March 2006) ] How to deal with millions of images ? How to accumulate general knowledge automatically from the Web ? How to deal with billions of shared users’ perceptions stored at massive scale ? How to realize the vision of social search? How to realize a roboter out�of�the�box? K. Kersting 10 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 5

Machine Learning in an Exponential World Machine Learning = Data + Model [Fergus et al. PAMI 30(11) 2008; Halevy et al., IEEE Intelligent Systems, 24 2009] � Most effort has gone into the modeling part � How much can the data itself help us to solve a problem? 11 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Spectrally Hashed Logistic Regression [Behley, K, Schulz, Steinhage, Cremers IROS10] Classes: Car Vegetation Ground Building Ground Truth Our Approach K. Kersting 12 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 6

Machine Learning in an Exponential World Machine Learning = Data + Model ML = Structured Data + Model + Reasoning � Real world is structured in terms of objects and relations � Relational knowledge can reveal additional correlations between variables of interest . Abstraction allows one to compactly model general knowledge and to move to complex inference [Fergus et al. PAMI 30(11) 2008; Halevy et al., IEEE Intelligent Systems, 24 2009] � Most effort has gone into the modeling part � How much can the data itself help us to solve a problem? 13 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 http://www.cs.washington.edu/research/textrunner/ [Etzioni et al. ACL08] Object Relation Uncertainty Object So, how do computer No complex inference (yet) ! systems deal with uncertainty, objects, and TextRunner : (Turing, born in, London) relations? How do we + WordNet : (London, part of, England) realize the vision of a + Rule : ‘born in’ is transitive thru ‘part of’ world:wide:mind? Conclusion: (Turing, born in, ������� ) “Programs will consume, combine, and correlate everything in the K. Kersting 14 universe of structured information and help users reason over it.” Lifted Message Passing GRAPHBOT@ ROS 2010 [S. Parastatidis et al., CACM Vol. 52(12):33�37 ] Taipei, Taiwan, October 22, 2010 7

(First:order) Logic handles Complexity E.g., rules of chess (which is a tiny problem): 1 page in first�order logic, daugther�of(cecily,john) ~100000 pages in propositional logic, daugther�of(lily,tom) ~100000000000000000000000000000000000000 pages as atomic�state model M � Many types of entities � Relations between them Explicit enumeration � Arbitrary knowledge 19 th C 5 th C B.C. Logic true/false atomic propositional first�order/relational 15 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Probability handles Uncertainty 17 th C 20 th C Probability � Sensor noise � Human error � Inconsistencies � Many types of entities � Unpredictability � Relations between them Explicit enumeration � Arbitrary knowledge 5 th C B.C. 19 th C Logic true/false atomic propositional first�order/relational K. Kersting 16 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 8

Will Traditional AI Scale ? “Scaling up the environment will inevitably overtax the resources of the current AI architecture.” 17 th C 20 th C Probability � Sensor noise � Human error � Inconsistencies � Many types of entities � Unpredictability � Relations between them Explicit enumeration � Arbitrary knowledge 19 th C 5 th C B.C. Logic true/false atomic propositional first�order/relational 17 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Statistical Relational Learning / AI (StarAI*) Let‘s deal with uncertainty, objects, and Natural domain modeling: � objects, properties, relations jointly relations A Compact, natural models � Probability Properties of entities can Planning � depend on properties of SAT Statistics related entities Logic Generalization over a � variety of situations Graphs Trees Learning The study and design of Search CV intelligent agents that act in Robotics noisy worlds composed of M unifies logical and statistical AI, objects and relations among M solid formal foundations, the objects M is of interest to many communities. K. Kersting 18 (*)First StarAI workshop at AAAI10;co�chaired with S. Russell, L. Kaelbling, A.Halevy, S. Natarajan, and L. Milhalkova Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 9

Pros of SRL / StarAI Relations can reveal additional +++ correlations. Abstraction allows for generalization and compactness � Better performance � Better understanding of domains SRL/StarAI techniques � Growth path for machine learning, have the potential to lay artificial intelligence, and robotics the foundations of next generation AI systems 19 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 For example, we can A M learn probabilistic relational models automatically from � millions of inter�related objects M generate optimal plans and learn to act optimally in uncertain � environments involving millions of objects and relations among them K. Kersting 20 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 10

[Richardson, Domingos MLJ 62(1�2): 107�136, 2006] Markov Logic Networks Suppose we have constants: alice , bob and p1 ∀ ∧ ⇒ x author ( x , p ) smart ( x ) high _ quality ( p ) 1 . 5 ∀ ⇒ x high _ quality ( p ) accepted ( p ) 1 . 1 ( ) ∀ ⇒ ⇔ x , y co _ author ( x , y ) smart ( x ) smart ( y ) 1 . 2 ∞ ∀ ∃ ∧ ⇒ x , y p author ( x , p ) author ( y , p ) co _ author ( x , y ) co_author(bob,alice) co_author(alice,bob) co_author(alice,alice) co_author(bob,bob) author(p1,alice) smart(alice) smart(bob) author(p1,bob) high_quality(p1) accepted(p1) 21 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Relational Exploration [Lange, Toussaint, K ECML:PKDD10] Placeholders (X,Y A) allow one to generalize experience fully grounded network!! Simulated robot manipulation domain (realistic physics engine, > 2^100 states) � Robot starts from zero knowledge and actively generates training trajectories � Goal: pile objects K. Kersting 22 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 11

Pros and Cons of SRL / StarAI Relations can reveal additional +++ correlations. Abstraction allows for generalization and compactness � Better performance � Better understanding of domains SRL/StarAI techniques � Growth path for machine learning, have the potential to lay artificial intelligence, and robotics the foundations of next generation AI systems ::: � Learning is much harder � Inference becomes a crucial issue Yes, SRL/StarAI is challenging but knowing � Greater complexity for user one of its ingredients is half the battle 23 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 So, can we do better? Inference in first�order logic is not „ground“ � M it is lifted, i.e., it never „touches“ the ground � Yes, we can and it is exactly the focus of my talk talk Resulting lifted approaches are often • Faster • More compact ... and provide more structure for optimization K. Kersting 24 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 12

Distributions can naturally be represented as factor graphs Random variable Factor resp. unnormalized ! potential There is an edge between a circle and a box if the variable is in � the domain/scope of the factor 25 K. Kersting Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 Factor Graphs from Graphical Models K. Kersting 26 Lifted Message Passing GRAPHBOT@ ROS 2010 Taipei, Taiwan, October 22, 2010 13