 |
Center for Nonlinear Studies |
Learning good feature representation (rather than pixels) is a fundamental goal in computer vision. Many computer vision methods rely on the availability of labeled data to produce feature representations from inputs. While labeled data is very expensive to get and sometimes too scarce to fit a model in real-world application (e.g., high-dimensional video analysis), unlabeled data can often be obtained in large scale at very low cost. In this talk, I will describe a class of unsupervised learning methods to generate good internal representation from unlabeled data. The approach is based on a generalization of generative models with sparse constraints, which emphasizes feedback processes as generators of local image predictions in hierarchical architectures. The Bayesian framework is utilized to address visual inference in the hierarchical structure, where each cortical area is an expert for inferring certain aspects of the visual scene. The learned sparse internal representations show favorable performance in variety of vision tasks, including generic object recognition, object detection and segmentation, image denoising and compression and vision-based autonomous navigation. The general principle of unsupervised sparse learning can also be applied to other domains than vision, such as biomimetic oder discrimination, text document retrieval and classification, etc.
Host: Kipton Barros, T-4 and CNLS