OccamNets: Mitigating Dataset Bias by Favoring Simpler Hypotheses

TL;DR

Dataset bias and spurious correlations undermine generalization in deep networks. OccamNets introduce architectural inductive biases that favor simpler hypotheses by exiting early and restricting prediction to fewer spatial regions via suppressed CAMs, learned without explicit bias labels. Across BiasedMNIST, COCO, and BAR, OccamNets outperform standard architectures and rival debiasing methods, with further gains when combined with existing techniques; ablations confirm the necessity of both inductive biases. The approach generalizes across backbones, reduces reliance on spurious cues, and remains effective under distribution shifts, offering a practical, architecture-centered avenue for debiasing in vision systems.

Abstract

Dataset bias and spurious correlations can significantly impair generalization in deep neural networks. Many prior efforts have addressed this problem using either alternative loss functions or sampling strategies that focus on rare patterns. We propose a new direction: modifying the network architecture to impose inductive biases that make the network robust to dataset bias. Specifically, we propose OccamNets, which are biased to favor simpler solutions by design. OccamNets have two inductive biases. First, they are biased to use as little network depth as needed for an individual example. Second, they are biased toward using fewer image locations for prediction. While OccamNets are biased toward simpler hypotheses, they can learn more complex hypotheses if necessary. In experiments, OccamNets outperform or rival state-of-the-art methods run on architectures that do not incorporate these inductive biases. Furthermore, we demonstrate that when the state-of-the-art debiasing methods are combined with OccamNets results further improve.

Figures (9)

• Figure 1: OccamNets focus on architectural inductive biases, which is an orthogonal direction to tackling dataset biases compared to the existing works.
• Figure 2: OccamNets are multi-exit architectures capable of exiting early through the exit decision gates. The exits yield class activation maps that are trained to use a constrained set of visual regions.
• Figure 3: For each dataset, the first two columns show bias-aligned (majority) samples, and the last column shows bias-conflicting (minority) samples. For BAR, the train set does not contain any bias-conflicting samples.
• Figure 4: Percentage of samples exited (Exit%) from each exit (barring $E_0$).
• Figure 5: Original image, and Grad-CAM visualizations for ERM and PGI on ResNet, and CAM visualizations on OccamResNet. The visualizations are for the ground truth.