Solving the Multiple-Instance Problem with Axis-Parallel Rectangles By Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez Appeared in Artificial.

Slides:



Advertisements
Similar presentations
Artificial Intelligence
Advertisements

Viktor Zhumatiya, Faustino Gomeza,
An Introduction of Support Vector Machine
Support Vector Machines
Machine learning continued Image source:
Complexity ©D Moshkovitz 1 Approximation Algorithms Is Close Enough Good Enough?
Chapter 8 Linear Regression © 2010 Pearson Education 1.
Infinite Horizon Problems
Association Analysis. Association Rule Mining: Definition Given a set of records each of which contain some number of items from a given collection; –Produce.
Content Based Image Clustering and Image Retrieval Using Multiple Instance Learning Using Multiple Instance Learning Xin Chen Advisor: Chengcui Zhang Department.
Fall 2006CENG 7071 Algorithm Analysis. Fall 2006CENG 7072 Algorithmic Performance There are two aspects of algorithmic performance: Time Instructions.
Multiple Instance Learning
x – independent variable (input)
Co-Training and Expansion: Towards Bridging Theory and Practice Maria-Florina Balcan, Avrim Blum, Ke Yang Carnegie Mellon University, Computer Science.
Multiple-Instance Learning Paper 1: A Framework for Multiple-Instance Learning [Maron and Lozano-Perez, 1998] Paper 2: EM-DD: An Improved Multiple-Instance.
Active Learning with Support Vector Machines
Image Categorization by Learning and Reasoning with Regions Yixin Chen, University of New Orleans James Z. Wang, The Pennsylvania State University Published.
Marakas: Decision Support Systems, 2nd Edition © 2003, Prentice-Hall Chapter Chapter 7: Expert Systems and Artificial Intelligence Decision Support.
CHAPTER 2 ANALYSIS OF ALGORITHMS Part 2. 2 Running time of Basic operations Basic operations do not depend on the size of input, their running time is.
Semi-supervised protein classification using cluster kernels Jason Weston, Christina Leslie, Eugene Ie, Dengyong Zhou, Andre Elisseeff and William Stafford.
Region Based Image Annotation Through Multiple-Instance Learning By: Changbo Yang Wayne State University Department of Computer Science.
CPSC 322 Introduction to Artificial Intelligence December 1, 2004.
Wayne State University, 1/31/ Multiple-Instance Learning via Embedded Instance Selection Yixin Chen Department of Computer Science University of.
CSCI 5582 Fall 2006 CSCI 5582 Artificial Intelligence Lecture 22 Jim Martin.
Abstract A new Open Artwork System Interchange Standard (OASIS) has been recently proposed for replacing the GDSII format. A primary objective of the new.
Randomized Variable Elimination David J. Stracuzzi Paul E. Utgoff.
Heuristic Search Introduction to Artificial Intelligence COS302 Michael L. Littman Fall 2001.
August 16, 2015EECS, OSU1 Learning with Ambiguously Labeled Training Data Kshitij Judah Ph.D. student Advisor: Prof. Alan Fern Qualifier Oral Presentation.
PLANOMETRIC VIEW OF A KITCHEN.
A Geometric Framework for Unsupervised Anomaly Detection: Detecting Intrusions in Unlabeled Data Authors: Eleazar Eskin, Andrew Arnold, Michael Prerau,
 Optimal Packing of High- Precision Rectangles By Eric Huang & Richard E. Korf 25 th AAAI Conference, 2011 Florida Institute of Technology CSE 5694 Robotics.
Introduction to variable selection I Qi Yu. 2 Problems due to poor variable selection: Input dimension is too large; the curse of dimensionality problem.
1 SUPPORT VECTOR MACHINES İsmail GÜNEŞ. 2 What is SVM? A new generation learning system. A new generation learning system. Based on recent advances in.
Multiple Instance Real Boosting with Aggregation Functions Hossein Hajimirsadeghi and Greg Mori School of Computing Science Simon Fraser University International.
ALG0183 Algorithms & Data Structures Lecture 6 The maximum contiguous subsequence sum problem. 8/25/20091 ALG0183 Algorithms & Data Structures by Dr Andy.
Ch10 Machine Learning: Symbol-Based
NUS CS5247 Deadlock-Free and Collision-Free Coordination of Two Robot Manipulators By Patrick A. O’Donnell and Tomás Lozano-Pérez MIT Artificial Intelligence.
Artificial Intelligence in Game Design N-Grams and Decision Tree Learning.
1 Markov Decision Processes Infinite Horizon Problems Alan Fern * * Based in part on slides by Craig Boutilier and Daniel Weld.
Extending the Multi- Instance Problem to Model Instance Collaboration Anjali Koppal Advanced Machine Learning December 11, 2007.
Machine Learning Chapter 5. Artificial IntelligenceChapter 52 Learning 1. Rote learning rote( โรท ) n. วิถีทาง, ทางเดิน, วิธีการตามปกติ, (by rote จากความทรงจำ.
Linear Programming Erasmus Mobility Program (24Apr2012) Pollack Mihály Engineering Faculty (PMMK) University of Pécs João Miranda
CSE554Fairing and simplificationSlide 1 CSE 554 Lecture 6: Fairing and Simplification Fall 2012.
Artificial Intelligence in Game Design Complex Steering Behaviors and Combining Behaviors.
CS654: Digital Image Analysis
D. M. J. Tax and R. P. W. Duin. Presented by Mihajlo Grbovic Support Vector Data Description.
1 Markov Decision Processes Infinite Horizon Problems Alan Fern * * Based in part on slides by Craig Boutilier and Daniel Weld.
Chapter 11 Statistical Techniques. Data Warehouse and Data Mining Chapter 11 2 Chapter Objectives  Understand when linear regression is an appropriate.
Multiple Instance Learning via Successive Linear Programming Olvi Mangasarian Edward Wild University of Wisconsin-Madison.
Automated Reasoning Early AI explored how to automated several reasoning tasks – these were solved by what we might call weak problem solving methods as.
Multiple Instance Learning for Sparse Positive Bags Razvan C. Bunescu Machine Learning Group Department of Computer Sciences University of Texas at Austin.
Ensemble Methods in Machine Learning
The importance of a good representation Properties of a good representation: Reveals important features Hides irrelevant detail Exposes useful constraints.
Simplex Method Simplex: a linear-programming algorithm that can solve problems having more than two decision variables. The simplex technique involves.
CS 8751 ML & KDDComputational Learning Theory1 Notions of interest: efficiency, accuracy, complexity Probably, Approximately Correct (PAC) Learning Agnostic.
1 Chapter 8: Model Inference and Averaging Presented by Hui Fang.
 Your company has been selected to design a new animal cracker box.  Designs must be creative and eye catching, yet still meet the necessary requirements.
CHAN Siu Lung, Daniel CHAN Wai Kin, Ken CHOW Chin Hung, Victor KOON Ping Yin, Bob Fast Algorithms for Projected Clustering.
Computational Properties of Perceptron Networks n CS/PY 399 Lab Presentation # 3 n January 25, 2001 n Mount Union College.
Jump to first page Relational Data. Jump to first page Inductive Logic Programming (ILP) n Can use ILP to find a set of rules capturing a property that.
Approximating Set Cover
Challenges in Creating an Automated Protein Structure Metaserver
B. Jayalakshmi and Alok Singh 2015
K-means and Hierarchical Clustering
Learning.
Revision (Part II) Ke Chen
Heuristic Search Thank you Michael L. Littman, Princeton University for sharing these slides.
Revision (Part II) Ke Chen
Mathematical Analysis of Algorithms
Basics Prof. Hsin-Mu (Michael) Tsai (蔡欣穆)
Presentation transcript:

Solving the Multiple-Instance Problem with Axis-Parallel Rectangles By Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez Appeared in Artificial Intelligence, Volume 89, Issue1-2, (January 1997), Pages: 31 – 71. Presented by Shuiwang Ji Part of this slides are from the original paper and from Dr. Yixin Chen.

The key-lock problem Problem statements: 1.There is a keyed lock on the door to the supply room; 2.Each staff member has a key chain containing several keys; 3.One key on each key chain can open the supply room door; 4.For some staff members, their supply key may open one or more other doors; Positive bag Negative bag

The key-lock problem Question: You are a lock smith and you are attempting to infer the most general required shape that a key must have in order to open the supply room door.

Drug activity prediction Problem statements: A good drug molecule will bind very tightly to the desired binding site; The input object is a molecule, and the observed result is “bind” or “not bind”; Conformation determines bind or not; Each drug molecule has many conformations;

Drug activity prediction Problem statements: A good drug molecule will bind very tightly to the desired binding site; The input object is a molecule, and the observed result is “bind” or “not bind”; Conformation determines bind or not; Each drug molecule has many conformations; Positive bag Negative bag

Drug activity prediction Question: The goal is to infer the most general shape required for binding to the binding site.

Supervised learning

Multiple-instance learning Key chain Molecule Key Conformations Open? Bind? bag

Formal problem definition Supervised learning: Multiple-instance learning:

Ray-based representation A set of 162 rays emanating from the origin; Each feature value is the distance from the origin to the molecular surface.

Three classes of algorithms

Standard APR algorithm x1x1 x2x2 2 For each negative instance, count the number of instances that must be excluded from the APR in order to exclude the negative instance. Greedy algorithm: Eliminate the negative instance that requires eliminating the fewest positive instances

Standard APR algorithm Only the NUMBER of positive instances is considered when eliminating negative instances; The resulting APR may not cover at least one instance from all positive bags; The cost of eliminating each positive instance should be different.

Outside-in APR algorithm x1x1 x2x2 2 Consider excluding positive instances that are expendable in the sense that every positive molecule still has at least one positive instance Greedy algorithm: define a cost function for the elimination 4

Outside-in APR algorithm x1x1 x2x2 2 Consider excluding positive instances that are expendable in the sense that every positive molecule still has at least one positive instance Greedy algorithm: define a cost function for the elimination 4

Outside-in APR algorithm The cost of excluding a positive instance of molecule depends on the other not yet excluded positive instances of the same molecule; A cost function based on the density estimation of the positive instances was proposed; The employed density estimation is expensive to compute.

Inside-out APR algorithm x1x1 x2x2 2 Choose an initial seed positive instance Grow the APR by identifying the positive instance that when added to the APR would least increase its size.

Grow: An algorithm for growing an APR with tight bounds along a specified set of features Inside-out APR algorithm Discrimination: An algorithm for choosing a set of discriminating features Expand: An algorithm for expanding the bounds of an APR to improve its generalization ability

Inside-out APR algorithm (grow) The size of an APR is the sum of the widths of all of its bounds. Greedy: Identify the positive instance of a not yet covered positive molecule that would least increase its size; 1,2,…,d-1,d, d+1 Backfitting: After making the d-th decision, all previous d- 1 decisions are revisited until no changes. 1,2,…,d-1, d, d+1

Inside-out APR algorithm (discrimination) A feature strongly discriminates against a negative instance if that instance is far outside of the bound of this feature; In each step, choose a feature that strongly discriminates against the largest number of negative instances; Repeat until all negative instances are excluded.

Inside-out APR algorithm (expand) APR resulting from the first two steps is too tight since it is designed to cover at least one positive instance of each positive bag; Apply kernel density estimation to estimate the probability that a positive instance will satisfy the bounds on that feature; Expand the bounds so that with high probability, new positive instance will fall inside the APR.

Performance of iterated discrimination on artificial data set

Performance of iterated discrimination on musk data set 1

Performance of iterated discrimination on musk data set 2

Thank you!