IC-153ResNet50 relies on flower petals and green background features as shortcuts when classifying bee images

Tue Minh Cao, Nhat Hoang-Xuan, Hieu Pham, Phi Le Nguyen, My T. Thai

SourceNeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions

Using NeurFlow to identify neuron groups in layer 4.2 of ResNet50 for the 'bee' class, the authors find three concept groups: 'pink petals', 'green background', and 'bee'. Most false-positive images (non-bee images classified as bee) score high on the 'petal' and 'green background' groups. Masking all neurons in the petal or green-background groups drastically distorts the prediction probability away from 'bee', while masking an equal number of random neurons yields negligible changes. This causally confirms the flower-petal bias previously identified correlationaly by Kim et al. (2024) and additionally reveals a 'green background' bias not reported in that prior work.

Evidence
interventional
Caveat
The experiment is demonstrated on a single class ('bee') and a single layer (layer 4.2), though the authors state the result is 'consistent for multiple layers and with different classes.' No specific probability values are printed in the text.
Model
ResNet / ResNet-152 / ResNet-101 / ResNet-50-BN ResNet50
Concepts
Shortcut
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Methods
Integrated Gradients / Integral of gradients [supporting]
Extraction
automatic-extraction