| This thesis studies machine learning problems involved in visual recognition on a variety of computer vision tasks. It attacks the challenge of scaling-up learning to efficiently handle more training data in object recognition, more noise in brain activation patterns, and learning more capable visual similarity models.;For learning similarity models, one challenge is to capture from data the subtle correlations that preserve the notion of similarity relevant to the task. Most previous work focused on improving feature learning and metric learning separately. Instead, we propose a unified deep-learning modeling framework that jointly optimizes the two through back-propagation. We model the feature mapping using a convolutional neural network and the metric function using a multi-layer fully-connected network. Enabled by large datasets and a sampler to handle the intrinsic imbalance between positive and negative samples, we are able to learn such models efficiently. We apply this approach to patch-based image matching and cross-domain clothing-item matching.;For analyzing activation patterns in images acquired using functional Magnetic Resonance Imaging (fMRI), a technology widely used in neuroscience to study human brain, challenges are small number of examples and high level of noise. The common ways of increasing the signal to noise ratio include adding more repetitions, averaging trials, and analyzing statistics maps solved based on a general linear model. In collaboration with neuroscientists, we developed a machine learning approach that allows us to analyze individual trials directly. This approach uses multi-voxel patterns over regions of interest as feature representation, and helps discover effects previous analyses missed.;For multi-class object recognition, one challenge is learning a non-one-vs-all multi-class classifier with large numbers of categories each with large numbers of examples. A common approach is data parallelization in a synchronized fashion: evenly and randomly distribute the data into splits, learn a full model on each split and average the models. We reformulate the overall learning problem in a consensus optimization framework and propose a more principled synchronized approach to distributed training. Moreover, we develop an efficient algorithm for solving the sub-problem by reducing it to a standard problem with warm start. |
