[Paper Review] Learning from Failure: Training Debiased Classifier from Biased Classifier

Jun Hyun Nam, Hyuntak Cha+3 more|arXiv (Cornell University)|Jul 6, 2020

Adversarial Robustness in Machine LearningComputer Science28 references53 citations

TL;DR

LfF trains a biased and a debiased network in parallel; the biased network amplifies bias early, while the debiased network focuses on samples where the bias fails, improving unbiased accuracy without explicit bias labels.

ABSTRACT

Neural networks often learn to make predictions that overly rely on spurious correlation existing in the dataset, which causes the model to be biased. While previous work tackles this issue by using explicit labeling on the spuriously correlated attributes or presuming a particular bias type, we instead utilize a cheaper, yet generic form of human knowledge, which can be widely applicable to various types of bias. We first observe that neural networks learn to rely on the spurious correlation only when it is "easier" to learn than the desired knowledge, and such reliance is most prominent during the early phase of training. Based on the observations, we propose a failure-based debiasing scheme by training a pair of neural networks simultaneously. Our main idea is twofold; (a) we intentionally train the first network to be biased by repeatedly amplifying its "prejudice", and (b) we debias the training of the second network by focusing on samples that go against the prejudice of the biased network in (a). Extensive experiments demonstrate that our method significantly improves the training of the network against various types of biases in both synthetic and real-world datasets. Surprisingly, our framework even occasionally outperforms the debiasing methods requiring explicit supervision of the spuriously correlated attributes.

Motivation & Objective

• Motivate debiasing when explicit bias labels are expensive or unavailable.
• Investigate training dynamics showing bias is learned when it is easier than the target attribute.
• Develop a two-network debiasing scheme that uses bias-focused amplification and failure-driven reweighting.
• Evaluate LfF on synthetic and real-world biased datasets to demonstrate improved unbiased performance.

Proposed method

• Train two networks simultaneously: a biased model fB and a debiased model fD.
• Use generalized cross entropy (GCE) loss to amplify bias in fB, emphasizing easy samples.
• Compute a relative difficulty weight W(x) = CE(fB(x),y) / (CE(fB(x),y) + CE(fD(x),y)) to reweight CE loss for fD.
• Update fD by a weighted CE loss using W(x) to emphasize bias-conflicting samples.
• Rely on the observation that bias is learned earlier than target attributes in malignant biases to guide weighting.
• Optionally compare against baselines that rely on explicit bias supervision.

Experimental results

Research questions

• RQ1 Can debiasing be achieved without explicit bias labels by exploiting training dynamics where bias is learned first?
• RQ2 Does a pair of networks—one biased and one debiased—improve unbiased accuracy on biased datasets?
• RQ3 How does the proposed GCE-based bias amplification interact with debiasing via weighted loss on bias-conflicting samples?
• RQ4 Is the approach robust across synthetic biased datasets and real-world biased datasets like CelebA and BAR?

Key findings

• LfF significantly improves unbiased accuracy across datasets with color bias, texture bias, and gender bias.
• On Colored MNIST with high bias-aligned ratio, unbiased accuracy improves from 35.34% (vanilla) to 63.39% (LfF).
• On Corrupted CIFAR-10 1 with 99.5% bias-aligned samples, unbiased accuracy improves from 17.93% to 31.66% (LfF).
• On CelebA, LfF yields higher unbiased accuracy and competitive bias-conflicting performance compared to Group DRO (which uses explicit bias labels).
• In BAR real-world data, LfF outperforms vanilla and ReBias on most action classes, indicating effectiveness in realistic bias scenarios.