Wayne State University, 1/31/ Multiple-Instance Learning via Embedded Instance Selection Yixin Chen Department of Computer Science University of New Orleans

Wayne State University, 1/31/20062 Outline An overview of multiple-instance learning Multiple-instance learning via embedded instance selection Experimental results and discussions Conclusions and future work

Wayne State University, 1/31/20063 Supervised Learning Sample i Result i Unknown Process, Result i Fixed-length vector of attribute values (usually called a “feature vector”) Result = f (Sample) Classification problem: if Result is discrete or categorical Regression problem: if Result is continuous

Wayne State University, 1/31/20064 Multiple-Instance Learning Problem Sample i, Result i Instance 2 Instance 1 Instance n Instance 3 =Result i Unknown Process Each instance is a fixed-length feature vector discrete or continuous Label i A label is associated with a bag, not the instances in the bag Bag i

Wayne State University, 1/31/20065 Drug Activity Prediction Predict whether a candidate drug molecule will bind strongly to a target protein Binding strength is largely determined by the shape of drug molecules Different conformations a Butane molecule (C 4 H 10 ) can take on. The molecule can rotate about the bond between the two central carbon atoms. (© 1998 by Oded Maron) Bag: a molecule Instance: a shape conformation Goal: predict whether a molecule binds to a protein of interest, and find the conformation that binds

Wayne State University, 1/31/20066 Object Recognition Bag: an image Instance: a salient region in an image Goal: predict whether an image contains an object, and identify the regions corresponding to the object

Wayne State University, 1/31/20067 Multiple-Instance Learning Models A bag is positive if and only if it contains at least one positive instance x1x1 x2x2 Axis-Parallel Rectangles Algorithm (APR) [Dietterich, et al., AI 1997] Minimal APR * * * * * * o o o o o o o o * * * * * * o o o o o oo o But there may not exist an APR that contains at least one instance from each positive bag and no instance from any negative bags

Wayne State University, 1/31/20068 Multiple-Instance Learning Models x1x1 x2x2 Diverse Density Algorithm (DD) [Maron and Lozano-Pérez, NIPS 1998] * * * * * * o o o o o o o o * * * * * * o o o o o oo o The diverse density at a location is high if the location is close to instances from different positive bags and is far way from all instances in negative bags Searching for an “axis-parallel ellipse” with high diverse density Sensitive to noise High computational cost.

Wayne State University, 1/31/20069 Multiple-Instance Learning Models x1x1 x2x2 EM-DD Algorithm [Zhang and Goldman, NIPS 2001] * * * * * * o o o o o o o o * * * * * * o o o o o oo o The diverse density is approximated by the “most likely” instance in each bag Finding an “axis-parallel ellipse” with high diverse density Sensitive to noise Cannot learn complex concepts.. nose mouth eyes

Wayne State University, 1/31/ Multiple-Instance Learning Models x1x1 x2x2 DD-SVM Algorithm [Chen and Wang, JMLR 2004] o o o o o o o o.. z1z1 z2z2 z1z1 z2z2 o 1 * * * * * * * Sensitive to noise Computational cost Instance classification Instance Prototype 1 Instance Prototype 2

Wayne State University, 1/31/ Outline An overview of multiple-instance learning Multiple-instance learning via embedded instance selection

Wayne State University, 1/31/ Motivation A bag is positive if it contains instances from at least two different distributions among N 1, N 2, and N 3 xkxk xixi xjxj N1N1 N3N3 N2N2 20 positive bags, 20 negative bags

Wayne State University, 1/31/ Motivation Embedding of bags Bags can be separated by a hyperplane Find the “right” embedding and the classifier 20 positive bags and 20 negative bags in the new feature space

Wayne State University, 1/31/ MILES: Multiple-Instance Learning via Embedded Instance Selection Instance-based feature mapping Joint feature selection and classification Minimizing a regularized training error 1-norm of w Hinge loss function 1-norm SVM

Wayne State University, 1/31/ MILES: Multiple-Instance Learning via Embedded Instance Selection Instance classification

Wayne State University, 1/31/ Outline An overview of multiple-instance learning Multiple-instance learning via embedded instance selection Experimental results and discussions

Wayne State University, 1/31/ Drug Activity Prediction MUSK1 and MUSK2 benchmark data sets A bag represents a molecule An instance represents a low-energy conformation of the molecule (166 features) Musk 1 Musk 2 # of bags# of instances/ bags# of positive bags

Wayne State University, 1/31/ Prediction Accuracy

Wayne State University, 1/31/ Region-Based Image Categorization COREL data set, 20 image categories, each containing 100 images AfricaBeachDogsWaterfall BuildingsBusesLizardAntiques DinosaursElephantsFashionBattle ships FlowersHorsesSunsetsSkiing MountainsFoodCarsDessert

Wayne State University, 1/31/ Sample Images

Wayne State University, 1/31/ Confusion Matrix

Wayne State University, 1/31/ Misclassified Images

Wayne State University, 1/31/ Performance Comparison Average classification accuracy

Wayne State University, 1/31/ Sensitivity to Labeling Noise

Wayne State University, 1/31/ Object Class Recognition Caltech data set Airplanes (800 images) Cars (800 images) Faces (435 images) Motorbikes (800 images) Background (2270 images) Salient region detector [Kadir and Brady, IJCV, 2001]

Wayne State University, 1/31/ Sample Images

Wayne State University, 1/31/ Performance Comparison True positive rate at the equal-error-rates point on the ROC curve

Wayne State University, 1/31/ Selected Features Patches related to positive features for object class ‘Airplanes’. Among 6821 features, 96 were selected as positive; 97 were selected as negative; 16 features were false positive. Patches related to positive features for object class ‘Cars’. Among features, 97 were selected as positive; 98 were selected as negative; 9 features were false positive.

Wayne State University, 1/31/ Selected Features Patches related to positive features for object class ‘Faces’. Among 6997 features, 42 were selected as positive; 27 were selected as negative; 14 features were false positive. Patches related to positive features for object class ‘Motorbikes’. Among 9995 features, 101 were selected as positive; 90 were selected as negative; 3 features were false positive.

Wayne State University, 1/31/ Instance Classification

Wayne State University, 1/31/ Instance Classification

Wayne State University, 1/31/ Computation Time Training time Over 10 folds500 imagesCars

Wayne State University, 1/31/ Outline An overview of multiple-instance learning Multiple-instance learning via embedded instance selection Experimental results and discussions Conclusions and future work

Wayne State University, 1/31/ Summary MILES Instance-based feature mapping Joint feature selection and classification using 1-norm SVM Instance classification Competitive performance in terms of accuracy, speed, and robustness to labeling noise

Wayne State University, 1/31/ Future Work Storage requirement A data matrix of size Sparseness Constraints on instances Model the spatial relationship among instances Learning parts in a 1-class setting

Wayne State University, 1/31/ Supported by Louisiana Board of Regents RCS Grant NASA EPSCoR DART Grant NSF EPSCoR Pilot Fund University of New Orleans The Research Institute for Children

Wayne State University, 1/31/ Acknowledgement Jinbo Bi, Siemens Medical Solutions Ya Zhang, University of Kansas Timor Kadir Rob Fergus

Wayne State University, 1/31/ More Information Papers in PDF, demonstrations, data sets, etc.