A0638
Title: Pairwise learning with deep neural networks
Authors: Junyu Zhou - University of Sydney (Australia) [presenting]
Abstract: Pairwise learning refers to learning tasks where the loss function considers a pair of samples simultaneously. Specific examples include ranking, pairwise least squares regression, and metric learning. The generalization analysis of pairwise learning with general losses is delved into by leveraging the specific structure of the target functions. Specifically, the target function of pairwise learning is demonstrated to exhibit (anti)symmetry if the loss function is (anti)symmetric. Building on this observation, structured (anti)symmetric deep neural networks are constructed to approximate the target function. Considering the hypothesis space consisting of these structured deep neural networks, an oracle-type inequality of the empirical minimizer is developed for pairwise learning. These results are applied to concrete examples such as ranking, pairwise least squares regression, and metric learning for illustration.
