Simultaneous Learning and Alignment: Multi-Instance and Multi-Pose Learning

IGERT 2 Electrical Engineering, California Institute of Technology [email protected] 3 Lab of Neuro Imaging University of California, Los Angeles [email protected] { } { } { } In object recognition in general and in face detection in particular, data alignment is necessary to achieve good classification results with certain statistical learning approaches such as Viola-Jones. Data can be aligned in one of two ways: (1) by separating the data into coherent groups and training separate classifiers for each; (2) by adjusting training samples so they lie in correspondence. If done manually, both procedures are labor intensive and can significantly add to the cost of labeling. In this paper we present a unified boosting framework for simultaneous learning and alignment. We present a novel boosting algorithm for Multiple Pose Learning (mpl), where the goal is to simultaneously split data into groups and train classifiers for each. We also bag labels: • Bag labels defined as but instance labels are unknown during training (latent variables). MIL-BOOST • Define bag probability as a softmax of instance probabilities: • Derivative of the loss function gives us the instance weights: • Algorithm Summary: • Instance labels are defined as where is a latent (unknown) variable, which is positive if example i is in group k. MPL-BOOST • Define probability as a softmax of probabilities determined by each classifier: • Derivative of the loss function gives us the instance weights for each classifier: • Algorithm Summary: • Boosting trains a strong classifier of the form • For a given loss function , perform gradient descent in function space. Each step is one weak classifier. • Let , and . • At step t we want a weak classifier that is close to the gradient: Gradient Boosting Review chosen weak classifier other weak classifiers current strong classifier = [ Re-derivation of Viola et al. 2005] review Multiple Instance Learning (mil), and in particular mil-boost, and describe how to use it to simultaneously train a classifier and bring data into correspondence. We show results on variations of LFW and MNIST, demonstrating the potential of these approaches. Experiments Experiments • Log likelihood is a standard loss function that we will use: