group norm for learning latent structural svms
DESCRIPTION
Group Norm for Learning Latent Structural SVMs Daozheng Chen (UMD, College Park), Dhruv Batra (TTI-Chicago), Bill Freeman (MIT), Micah K. Johnson (GelSight, Inc.). Overview. Induce Group Norm. Our goal Estimate model parameters Learn the complexity of latent variable space. - PowerPoint PPT PresentationTRANSCRIPT
Group Norm for Learning Latent Structural SVMsDaozheng Chen (UMD, College Park), Dhruv Batra (TTI-Chicago),
Bill Freeman (MIT), Micah K. Johnson (GelSight, Inc.)
Overview
• Data with complete annotation is rarely ever available.
• Latent variable models capture interaction betweeno observed data (e.g. gradient histogram image features)o latent or hidden variables not observed in the training data (e.g. location of object parts).
• Parameter estimation involve a difficult non-convex optimization problem (EM, CCCP, self-paced learning)
• Our goal• Estimate model parameters• Learn the complexity of latent variable space.
• Our approach• norm for regularization to estimate the parameters of a latent-variable model.
Latent Structural SVM
Prediction Rule:
Label space Latent Space
Joint feature vector
Inducing Group Normw partitioned into P groups; each group corresponds to the parameters of a latent variable state
Induce Group Norm
Alternating Coordinate and Subgradient Descent
nonconvex
convex
convex
Rewrite Learning Objective
Minimize Upper bound of
convex if {hi} is fixed
Digit recognition experiment (following the setup of Kumar et al. NIPS ‘10)• MNIST data: binary classification on four difficult digit pairs
• (1,7), (2,7), (3,8), (8,9)• Training data 5,851 - 6,742, and testing data 974 - 1,135 • Rotate digit images with angles from -60o to 60o • PCA to form 10 dimensional feature vector
Experiment
• Significantly higher accuracy than random sampling.
• 66% faster than full model with no loss in accuracy!
Digit Recognition
Key Contribution
-60o -48o -36o -24o -12o 0o 12o 24o 36o 48o 60o
-60o-48o -12o-60o -48o0o -48o-36o
l2 norm of the parameter vectors for different angles over the 4 digit pairs.• Select only a few angles, much fewer than 22 angles Angles Not Selected
ImagesRotation
(Latent Var.)Feature Vector
• At group level, the norm behave like norm and induces group sparsity.
norm for regularization
• Within each group, the norm behave like norm and does not promote sparsity.
Learning objective:
Subgradient
Felzenszwalb et al. car model on the PASCAL VOC 2007 data. Each row is a component of the model.
Root filters Part filters Part displacement
Component #1
Component #2