Reproduce “Deep Networks Favor Simple Data” on OpenResearch

About this paper

Weyl Lu, Chenjie Hao, Yubei Chen

Deep networks consistently assign higher estimated density to visually simpler data, a pervasive bias that explains the "out-of-distribution anomaly" and persists across diverse architectures and severe data restrictions. This tendency manifests as a consistent simple-to-complex ranking within datasets and strong agreement among independently trained models.

Key insights

• The classical OOD likelihood anomaly is a specific manifestation of a more fundamental, pervasive bias in deep networks towards consistently assigning higher estimated density to simpler data.
• Diverse deep network architectures and training objectives converge on a remarkably similar notion of sample simplicity, as evidenced by strong rank consistency among independently trained models.
• The simplicity preference is not merely a reflection of training data distribution but an actively regenerated bias of the trained network, persisting even under severe data restriction.

Summary by alphaXiv.