Reproduce “Concept Whitening for Interpretable Image Recognition” on OpenResearch

About this paper

Zhi Chen, Yijie Bei, Cynthia Rudin

Researchers at Duke University, including Cynthia Rudin, developed Concept Whitening, a trainable module that modifies neural networks during training to align their latent space dimensions with human-understandable concepts. This approach creates inherently interpretable models without compromising predictive accuracy, quantitatively outperforming existing post-hoc explanation methods in concept purity and separability.

Key insights

• Interpretability can be directly integrated into deep neural network architectures by shaping their latent spaces during training, rather than relying on post-hoc explanations.
• By decorrelating latent features and aligning them with human-defined concepts, Concept Whitening enables purer, more disentangled concept representations within the network.
• The method allows for observing how the network's understanding of concepts evolves across different layers, from primitive features to complex semantic representations.

Summary by alphaXiv.