Filter-decomposed Convolution in Deep Neural Networks: On Groups, Graphs, and Across Domains
Deep convolutional neural networks (CNN) have been developed and applied to data on Euclidean domains as well as non-Euclidean ones. In this talk, we introduce a framework of decomposing convolutional filters over a truncated set of basis filters, which applies to the standard CNN, group-equivariant CNN, as well as convolution on graphs. The basis decomposition reduces the model and computational complexity of deep CNNs with an automatically imposed filter regularity. First, for group equivariant CNNs, a joint basis decomposition over space and group geometry achieves group equivariance in image data, including rotation and scaling groups, with provable representation stability with respect to the geometric deformation of input data. Second, the decomposed convolution on graphs provides a unified framework for several graph convolution models. The graph convolution with low-rank local filters has enlarged expressiveness to represent graph signals than spectral graph convolutions, and shows empirical advantage on facial expression and action recognition datasets. At last, when allowing the light-weighted basis layer to be adapted to varying modals in data, the decomposition also provides a new way of invariant feature learning across domains, as well as conditional image generation. Joint work with Qiang Qiu, Ze Wang, Zichen Miao, Wei Zhu, Robert Calderbank, and Guillermo Sapiro.
Xiuyuan Cheng is an assistant professor of mathematics at Duke University. Dr. Cheng is interested in theoretical and computational problems in high-dimensional data analysis and machine learning, particularly on spectral methods, kernel matrices, and neural networks. Dr. Cheng's work is supported by NSF, NIH, and the Alfred P. Sloan Foundation.