[Paper Review] Learning De-biased Representations with Biased Representations

Hyojin Bahng, Sanghyuk Chun+3 more|arXiv (Cornell University)|Oct 7, 2019

Machine Learning and Data ClassificationComputer Science100 citations

TL;DR

The paper proposes ReBias, a regulariser that enforces learned representations to be statistically independent from a set of bias-driven representations, improving cross-bias generalisation without requiring explicit bias labels.

ABSTRACT

Many machine learning algorithms are trained and evaluated by splitting data from a single source into training and test sets. While such focus on in-distribution learning scenarios has led to interesting advancement, it has not been able to tell if models are relying on dataset biases as shortcuts for successful prediction (e.g., using snow cues for recognising snowmobiles), resulting in biased models that fail to generalise when the bias shifts to a different class. The cross-bias generalisation problem has been addressed by de-biasing training data through augmentation or re-sampling, which are often prohibitive due to the data collection cost (e.g., collecting images of a snowmobile on a desert) and the difficulty of quantifying or expressing biases in the first place. In this work, we propose a novel framework to train a de-biased representation by encouraging it to be different from a set of representations that are biased by design. This tactic is feasible in many scenarios where it is much easier to define a set of biased representations than to define and quantify bias. We demonstrate the efficacy of our method across a variety of synthetic and real-world biases; our experiments show that the method discourages models from taking bias shortcuts, resulting in improved generalisation. Source code is available at https://github.com/clovaai/rebias.

Motivation & Objective

• Define cross-bias generalisation and motivate why models rely on shortcut biases in training data.
• Introduce bias-characterising model class G to capture biases without requiring explicit bias labels.
• Propose the ReBias regulariser using HSIC to enforce independence between the main predictor and biased predictors.
• Provide a learning objective that jointly optimises task loss and de-biasing via a max-HSIC/minimisation framework.
• Demonstrate the method on synthetic and real-world biases (Biased MNIST, ImageNet, action recognition).

Proposed method

• Define a bias-characterising model class G that captures bias by design (e.g., small receptive field CNNs for texture bias, 2D CNNs for static bias).
• Formulate HSIC-based independence between the main predictor f(X) and biased predictors g(X) for g in G.
• Use a minimax-like objective to push f away from the biased predictions: minimize L(f,X,Y) + lambda * max_{g in G} HSIC(f(X), g(X)) with an additional term to regularise biased models: max_{g in G} (HSIC(f,g) - lambda_g L(g)).
• Train f and g alternately to ensure f learns signals S rather than biases B.
• Apply ReBias to settings where explicit bias labels are unavailable but biases can be captured by G, enabling de-biasing without data augmentation or explicit bias quantification.

Experimental results

Research questions

• RQ1 Can a de-biasing objective be formulated without explicit bias labels by using a bias-characterising model class G?
• RQ2 Does encouraging independence between the main model and biased models improve cross-bias generalisation across texture, static, and other biases?
• RQ3 How does HSIC-based regularisation influence the learned invariances of f relative to biased counterparts?
• RQ4 Is ReBias effective on synthetic (biased MNIST) and real-world biases (ImageNet, action recognition) compared to prior de-biasing methods?

Key findings

• ReBias improves unbiased accuracy across highly biased training scenarios in Biased MNIST, increasing accuracy from baseline biased performance to substantially higher unbiased performance.
• Compared to HEX and RUBi, ReBias achieves more uniform improvements across bias-class pairs and better unbiased generalisation in experiments.
• In ImageNet experiments, ReBias demonstrates improved cross-bias robustness relative to baseline and some prior approaches, as measured by texture-unbiased metrics and auxiliary benchmarks (ImageNet-A, ImageNet-C).
• Learning curves show a strong correlation between decreasing HSIC and improving unbiased accuracy, indicating the regulariser effectively promotes bias-robust representations.
• The method does not require explicit bias labels and relies on a bias-characterising model class G to capture plausible shortcuts, making it applicable across vision tasks like action recognition and image classification.