Learning Robust Representations by Projecting Superficial Statistics Out

Abstract

Despite impressive performance as evaluated on i.i.d. holdout data, deep neural networks depend heavily on superficial statistics of the training data and are liable to break under distribution shift. For example, subtle changes to the background or texture of an image can break a seemingly powerful classifier. Building on previous work on domain generalization, we hope to produce a classifier that will generalize to previously unseen domains, even when domain identifiers are not available during training. This setting is challenging because the model may extract many distribution-specific (superficial) signals together with distribution-agnostic (semantic) signals. To overcome this challenge, we incorporate the gray-level co-occurrence matrix (GLCM) to extract patterns that our prior knowledge suggests are superficial: they are sensitive to the texture but unable to capture the gestalt of an image. Then we introduce two techniques for improving our networks' out-of-sample performance. The first method is built on the reverse gradient method that pushes our model to learn representations from which the GLCM representation is not predictable. The second method is built on the independence introduced by projecting the model's representation onto the subspace orthogonal to GLCM representation's. We test our method on the battery of standard domain generalization data sets and, interestingly, achieve comparable or better performance as compared to other domain generalization methods that explicitly require samples from the target distribution for training.

To appear at ICLR 2019. Implementation: https://github.com/HaohanWang/HEX

Cited by in corpus (50)

• Learning Robust Global Representations by Penalizing Local Predictive Power
• Out-of-Distribution Generalization via Risk Extrapolation (REx)
• Domain Generalization by Marginal Transfer Learning
• Improving Robustness Without Sacrificing Accuracy with Patch Gaussian Augmentation
• The Pitfalls of Simplicity Bias in Neural Networks
• Domain Generalization using Causal Matching
• Discriminative Adversarial Domain Generalization with Meta-learning based Cross-domain Validation
• Maximum-Entropy Adversarial Data Augmentation for Improved Generalization and Robustness
• Efficient Domain Generalization via Common-Specific Low-Rank Decomposition
• Learning from Failure: Training Debiased Classifier from Biased Classifier
• Inspect, Understand, Overcome: A Survey of Practical Methods for AI Safety
• Learning Debiased Representation via Disentangled Feature Augmentation
• High Frequency Component Helps Explain the Generalization of Convolutional Neural Networks
• Self-Challenging Improves Cross-Domain Generalization
• Nonlinear Invariant Risk Minimization: A Causal Approach
• Selecting Data Augmentation for Simulating Interventions
• Unlearn Dataset Bias in Natural Language Inference by Fitting the Residual
• Robust and Generalizable Visual Representation Learning via Random Convolutions
• Can Subnetwork Structure be the Key to Out-of-Distribution Generalization?
• POEM: Polarization of Embeddings for Domain-Invariant Representations
• Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU Models
• Domain Invariant Representation Learning with Domain Density Transformations
• Generalizing Graph Neural Networks on Out-Of-Distribution Graphs
• Split Batch Normalization: Improving Semi-Supervised Learning under Domain Shift
• Proactive Pseudo-Intervention: Causally Informed Contrastive Learning For Interpretable Vision Models
• EnD: Entangling and Disentangling deep representations for bias correction
• Learn to Expect the Unexpected: Probably Approximately Correct Domain Generalization
• Model-Based Domain Generalization
• Learning Debiased and Disentangled Representations for Semantic Segmentation
• Self-training Avoids Using Spurious Features Under Domain Shift
• Toward Learning Human-aligned Cross-domain Robust Models by Countering Misaligned Features
• TRIER: Template-Guided Neural Networks for Robust and Interpretable Sleep Stage Identification from EEG Recordings
• A Bit More Bayesian: Domain-Invariant Learning with Uncertainty
• BiaSwap: Removing dataset bias with bias-tailored swapping augmentation
• Learning to Diversify for Single Domain Generalization
• Feature Stylization and Domain-aware Contrastive Learning for Domain Generalization
• Towards Robustifying NLI Models Against Lexical Dataset Biases
• StyleAugment: Learning Texture De-biased Representations by Style Augmentation without Pre-defined Textures
• Hierarchical Variational Auto-Encoding for Unsupervised Domain Generalization
• Robust Representation Learning via Perceptual Similarity Metrics
• Explainability-aided Domain Generalization for Image Classification
• Unsupervised Learning of Debiased Representations with Pseudo-Attributes
• Robust Domain-Free Domain Generalization with Class-aware Alignment
• Fighting Fire with Fire: Contrastive Debiasing without Bias-free Data via Generative Bias-transformation
• Reappraising Domain Generalization in Neural Networks
• Combating Unknown Bias with Effective Bias-Conflicting Scoring and Gradient Alignment
• Discriminative Domain-Invariant Adversarial Network for Deep Domain Generalization
• RC2020 Report: Learning De-biased Representations with Biased Representations
• Generalization in NLI: Ways (Not) To Go Beyond Simple Heuristics
• Neuron Coverage-Guided Domain Generalization