1. Sharing Features Between Visual Tasks at Different Levels of Granularity Sung Ju Hwang 1, Fei Sha 2 and Kristen Grauman 1 1 University of Texas at Austin, 2 University of Southern California Problem Sharing features between sub/superclasses A single visual instance can have multiple labels at different levels of semantic granularity.. Main Idea We propose to simultaneously learn shared features that are discriminative for tasks at different levels of semantic granularity Baselines Baselines Dataset Dataset Sharing features between objects/attributes Example object class / attribute predictions Example object class / attribute predictions 1) No sharing : Baseline SVM classifier for the object class recognition 2) Sharing-Same level only : Sharing features between object classifiers at the same level 3) Sharing+Superclass : Subclass classifiers trained with features shared with superclasses* 4) Sharing+Subclass : Superclass classifiers trained with features shared with subclasses* *We use the algorithm for kernel classifiers. 1) Finer grained categorization tasks benefit from sharing features with their superclasses. → A subclass classifier learns features specific to its superclass, so that it can discriminate better between itself and the classes that do NOT belong to the same superclass. 2) Coarser grained categorization tasks do not benefit from sharing features with their subclasses → Features learned for subclasses are just intra-class variances that introduce confusion. Recognition accuracy Recognition accuracy Predicted Object NSO NSA Ours Dolphin Walrus Grizzly bear NSO NSA Ours Grizzly bear Rhinoceros Moose NSO NSA Ours Giant Panda Rabbit Rhinoceros Fast, active, toughskin, chewteeth, forest, ocean, swims NSA Ours Fast, active, toughskin, fish, forest, meatteeth, strong Strong, inactive, vegetation, quadrapedal, slow, walks, big NSA Ours Strong, toughskin, slow, walks, vegetation, quadrapedal, inactive Quadrapedal, oldworld, walks, ground, furry, gray, chewteeth NSA Ours Quadrapedal, oldworld, ground, walks, tail, gray, furry White, Spots, Long leg Polar bear Motivation: By regularizing to use the shared features, we aim to select features that are associated with semantic concepts at each level avoid overfitting when object- labeled data is lacking 1) Our method is more robust to background clutter, as it has a more refined set of features from sparsity regularization with attributes. 2) Our method makes robust predictions in atypical cases. 3) When our method fails, it often makes more semantically “close” predictions. Visual instance Dalmatian Canine Spots Object class Dalmatian Attributes → How can we learn new information from these extra labels, that can aid object recognition? u2u2 u1u1 u3u3 uDuD x2x2 x1x1 x3x3 xDxD Shared features Input visual features Superclass Dog, Canine, Placental mammal Dataset# images# classes# AttributesHierarchy level Animals with Attributes30,47550(40)282 Recognition accuracy for each class Recognition accuracy for each class Overall recognition accuracy Overall recognition accuracy 1)We make improvement on 33 classes out of 50 AWA classes, and on all classes of OSR. 2)Classes with more distinct attributes benefit more from feature sharing: e.g. Dalmatian, leopard, giraffe, zebra 1) No sharing-Obj. :Baseline SVM classifier for object class recognition 2) No sharing-Attr. : Baseline object recognition on predicted attributes as in Lampert’s approach 3) Sharing-Obj. : Our multitask feature sharing with the object class classifiers only 4) Sharing-Attr. : Our multitask feature sharing method with object class + attribute classifiers Baselines Baselines Predicted Attributes Red: incorrect prediction Conclusion / Future Work By sharing features between classifiers learned at different levels of granularity, we improve object class recognition rates. The exploited semantics effectively regularize the object models. Future work 1) Automatic selection of useful attributes/superclass grouping. 2) Leveraging the label structure to refine degree of sharing. Algorithm Extension to Kernel classifiers Extension to Kernel classifiers 1) Formulate kernel matrix K 2) Compute the basis vector B and diagonal matrix S using Gram- Schmidt process 3) Transform the data according to the learned B and S Learning shared features for linear classifiers Learning shared features for linear classifiers 1) Initialize covariance matrix Ω with a scaled identity matrix I/D 2) Transform the variables using the covariance matrix Ω : transformed n-th feature vector : transformed classifier weights 3) Solve for the optimal weights, while holding Ω fixed. 4) Update the covariance matrix Ω Alternate until W converges : weight vectors : smoothing parameter (for numerical stability) Variable updates 4) Apply the algorithm for the linear classifiers on the transformed features Initialization Independent classifier training Feature learning We adopt the alternating optimization algorithm from [Argyriou08] that can train classifiers and learn the shared features at each step. Learning shared features via regularization Sharing features via sparsity regularization Sharing features via sparsity regularization Trace norm regularization. Multitask feature learning Multitask feature learning Object classifier Attribute classifier SVM loss function on the original feature space (2,1)-norm L2-norm: Joint data fitting L1-norm: sparsity sparsity general Training set specific : n-th feature vector : n-th label for task t : parameter (weight vector) for task t Multitask Feature Learning: Sparsity regularization on the parameters across different tasks results in shared features with better generalization power. : regularization parameter : Orthogonal transformation to a shared feature space Covariance matrix. [Argyriou08] M. Argyriou, T. Evgeniou, Convex Multi-task Feature Learning, Machine Learning, 2008 Transformed (Shared) features Sparsity regularizer loss function Convex optimization Convex optimization However, the (2,1)-norm is nonsmooth. We instead solve an equivalent form, in which the features are replaced with a covariance matrix that measures the relative effectiveness of each dimension. How can we promote common sparsity across different parameters? → We use a (2,1)-norm regularizer that minimizes L1 norm across tasks [Argyriou08]. Our approach makes substantial improvements over the baselines. Exploiting the external semantics the auxiliary attribute tasks provide, our learned features generalize better---particularly when less training data is available.