Improving deep neural network generalization and robustness to background bias via layer-wise relevance propagation...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Background bias in image datasets can lead to spurious correlations, causing shortcut learning in deep neural networks (DNNs).
• This phenomenon results in models performing well on training data but generalizing poorly to real-world scenarios.
• Examples include medical imaging, where background features in chest X-rays can bias classification for diseases like COVID-19 or tuberculosis.

Purpose of the Study:

• To develop a method that minimizes background bias and hinders shortcut learning in deep classifiers.
• To improve the generalization performance of DNNs on external datasets.
• To offer a computationally efficient and widely applicable solution to background bias.

Main Methods:

• The study proposes an approach (ISNet) that optimizes Layer-wise Relevance Propagation (LRP) heatmaps to mitigate background bias.
• Synthetic background bias was injected into image datasets for quantitative comparison.
• ISNet's performance was evaluated against eight state-of-the-art DNNs on biased and external datasets.

Main Results:

• ISNet demonstrated superior robustness against background bias compared to eight benchmark DNNs.
• The method significantly reduced shortcut learning by focusing on relevant image features, such as lungs in chest X-rays.
• ISNet achieved substantially better generalization performance on out-of-distribution test databases.

Conclusions:

• Optimizing LRP heatmaps is an effective strategy to combat background bias and shortcut learning in deep learning models.
• ISNet offers a light, fast, and broadly applicable solution without increasing computational costs.
• The proposed method enhances the reliability and real-world applicability of DNNs, particularly in critical domains like medical diagnostics.