Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EiHi Net: Out-of-Distribution Generalization Paradigm

Qinglai Wei, Beiming Yuan, Diancheng Chen

TL;DR

EiHi Net presents a two-paradigm OoD generalization framework that can be plugged into any visual backbone. The first paradigm is a supervised, pairwise decorrelation contrastive learning mechanism built around original–positive–negative minimal learning elements with a covariance regularizer to avoid collapse. The second paradigm uses a human-in-the-loop with a few guidance samples to prune background-sensitive feature dimensions, thereby suppressing spurious domain correlations. On the NICO dataset, EiHi Net achieves competitive OoD performance without domain labels or pretraining, and analyses show the covariance term is essential for robustness. The work also discusses limitations in style-based OoD settings and highlights the potential to reduce domain supervision in practical applications.

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 paradigm that can be blessed on any visual backbone. This paradigm can change the previous learning method of the deep model, namely find out correlations between inductive sample features and corresponding categories, which suffers from pseudo correlations between indecisive features and labels. We fuse SimCLR and VIC-Reg via explicitly and dynamically establishing the original - positive - negative sample pair as a minimal learning element, the deep model iteratively establishes a relationship close to the causal one between features and labels, while suppressing pseudo correlations. To further validate the proposed model, and strengthen the established causal relationships, we develop a human-in-the-loop strategy, with few guidance samples, to prune the representation space directly. Finally, it is shown that the developed EiHi net makes significant improvements in the most difficult and typical OoD dataset Nico, compared with the current SOTA results, without any domain ($e.g.$ background, irrelevant features) information.

EiHi Net: Out-of-Distribution Generalization Paradigm

TL;DR

EiHi Net presents a two-paradigm OoD generalization framework that can be plugged into any visual backbone. The first paradigm is a supervised, pairwise decorrelation contrastive learning mechanism built around original–positive–negative minimal learning elements with a covariance regularizer to avoid collapse. The second paradigm uses a human-in-the-loop with a few guidance samples to prune background-sensitive feature dimensions, thereby suppressing spurious domain correlations. On the NICO dataset, EiHi Net achieves competitive OoD performance without domain labels or pretraining, and analyses show the covariance term is essential for robustness. The work also discusses limitations in style-based OoD settings and highlights the potential to reduce domain supervision in practical applications.

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 paradigm that can be blessed on any visual backbone. This paradigm can change the previous learning method of the deep model, namely find out correlations between inductive sample features and corresponding categories, which suffers from pseudo correlations between indecisive features and labels. We fuse SimCLR and VIC-Reg via explicitly and dynamically establishing the original - positive - negative sample pair as a minimal learning element, the deep model iteratively establishes a relationship close to the causal one between features and labels, while suppressing pseudo correlations. To further validate the proposed model, and strengthen the established causal relationships, we develop a human-in-the-loop strategy, with few guidance samples, to prune the representation space directly. Finally, it is shown that the developed EiHi net makes significant improvements in the most difficult and typical OoD dataset Nico, compared with the current SOTA results, without any domain ( background, irrelevant features) information.
Paper Structure (11 sections, 7 equations, 7 figures, 10 tables, 1 algorithm)

This paper contains 11 sections, 7 equations, 7 figures, 10 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Description
  3. Related works
  4. Methdology
  5. Feature Extraction
  6. Spurious causal elimination
  7. EiHi Net
  8. Challenge
  9. Ablation Study
  10. Conclusion
  11. Style Information O.o.D.

Figures (7)

  • Figure 1: The figure shows the example of NICO-animal.
  • Figure 2: The figure shows the example of NICO-vehicle.
  • Figure 3: The figure shows the feedforward process of input data.
  • Figure 4: The domain offset factor leads to the generation of wrong causal chain.
  • Figure 5: The figure shows the process of eliminating the domain information dimension.
  • ...and 2 more figures