Pattern-set Representations using Linear, Shallow and Tensor Subspaces

Abstract

Pattern-set matching refers to a class of problems where learning takes place through sets rather than elements. Much used in computer vision, this approach presents robustness to variations such as illumination, intrinsic parameters of the signal capture devices, and pose of the analyzed object. Inspired by applications of subspace analysis, three new collections of methods are presented in this paper: (1) New representations for two-dimensional sets; (2) Shallow networks for image classification; and (3) Subspaces for tensor representation and classification. New representations are proposed with the aim of preserving the spatial structure and maintaining a fast processing time. We also introduce a technique to keep temporal structure, even using the principal component analysis, which classically does not model sequences. In shallow networks, we present two convolutional neural networks that do not require backpropagation, employing only subspaces for its convolution filters. These networks present advantages when the training time and hardware resources are scarce. Finally, to handle tensor data, such as video data, we propose methods that employ subspaces for representation in a compact and discriminative way. Our proposed work has been applied to several problems, such as 2D data representation, shallow networks for image classification, and tensor representation and learning.