Abstract
This paper develops a new EiHi net to solve the out-of-distribution (OoD)generalization problem in deep learning. EiHi net is a model learning paradigmthat can be blessed on any visual backbone. This paradigm can change theprevious learning method of the deep model, namely find out correlationsbetween inductive sample features and corresponding categories, which suffersfrom pseudo correlations between indecisive features and labels. We fuse SimCLRand VIC-Reg via explicitly and dynamically establishing the original - positive- negative sample pair as a minimal learning element, the deep modeliteratively establishes a relationship close to the causal one between featuresand labels, while suppressing pseudo correlations. To further validate theproposed model, and strengthen the established causal relationships, we developa human-in-the-loop strategy, with few guidance samples, to prune therepresentation space directly. Finally, it is shown that the developed EiHi netmakes significant improvements in the most difficult and typical OoD datasetNico, compared with the current SOTA results, without any domain ($e.g.$background, irrelevant features) information.