[Paper Review] Learning Robust Representations by Projecting Superficial Statistics Out

Haohan Wang, Zexue He+2 more|arXiv (Cornell University)|Mar 2, 2019

Domain Adaptation and Few-Shot LearningComputer Science93 citations

TL;DR

This paper introduces Neural Gray-Level Co-occurrence Matrix (NGLCM) to capture texture, and HEX to project out texture-related signals, yielding domain-generalization improvements without target-domain data.

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.

Motivation & Objective

• Motivate learning classifiers that generalize across unseen domains by reducing reliance on superficial statistics like texture/background.
• Develop a differentiable texture-only feature extractor (NGLCM) and a method (HEX) to discard texture information during learning.
• Demonstrate effectiveness on synthetic and standard domain generalization benchmarks without using target-domain samples during training.
• Assess how HEX compares to established DG methods across various datasets.

Proposed method

• Introduce Neural Gray-Level Co-occurrence Matrix (NGLCM) as a differentiable block capturing texture information while avoiding semantic content.
• Define G = s(a; φ_a) s^T(b; φ_b) where s is a clipped, differentiable thresholding function; map image pixels to a texture representation.
• Propose two HEX strategies: (i) Adversarially train a predictor to recover GLCM features from h(X; θ) and backpropagate to fool it (ADV/ADVE), (ii) Project F_A onto the orthogonal complement of F_G to obtain F_L ( HEX ).
• Use a two-branch architecture combining raw representation h(X; θ) and texture representation g(X; φ) to produce predictions, with testing time using the transformed representation F_L.
• Experiment with MNIST-family, synthetic nuisance-background facial expression data, MNIST rotation, and PACS to compare HEX/ADV against DG baselines such as DANN and InfoDropout.

Experimental results

Research questions

• RQ1 Can we encourage a model to rely less on superficial statistics without labeled target-domain data?
• RQ2 Do differentiable texture-based representations (NGLCM) and projection-based invariance (HEX) improve out-of-domain performance compared with existing DG methods?
• RQ3 How do HEX and NGLCM perform across synthetic and real domain-shift benchmarks (including PACS and MNIST-rotation)?
• RQ4 What are the trade-offs and stability considerations when jointly training NGLCM/HEX with the main classifier?

Key findings

• NGLCM primarily captures textural information and is not effective for semantic digit recognition, validating its texture-focused role.
• HEX improves robustness to various distribution shifts and often achieves competitive or superior performance compared with DG methods that require target-domain samples (e.g., DANN, Fusion) on several benchmarks.
• In MNIST-rotation experiments, HEX generally yields strong average performance, sometimes approaching or surpassing state-of-the-art domain generalization methods.
• On the PACS dataset, HEX is close to the Fusion method while using significantly fewer parameters, performing notably well on Art and Cartoon domains.
• Across synthetic nuisance-background tasks, ADV and HEX show stable improvements as domain correlation strength increases, with HEX providing complementary benefits to adversarial approaches.
• The authors discuss limitations, including incomplete removal of semantic information by NGLCM and potential training instability, mitigated by training heuristics.