Fig. 4: Diagnosis of the topological dataset’s shortcuts.
From: Mitigating data bias and ensuring reliable evaluation of AI models with shortcut hull learning
a Statistical characteristics of a 3 × 3 template in a 29 × 29-sized dataset. b The proportion of distinguishable classes under different template sizes in the 29 × 29 dataset. The horizontal axis represents the size of square templates, such as 1 × 1, 2 × 2, 3 × 3, etc. c Diagnosis based on the SHL by models with different inductive biases, including ResNet-5026, ViT-B/1641, RepVGG-A257, Swin-T58, PViG-S59, ResNeXt-5068, Inception-V363, ConvMixer-1024/1069, EfficientNet-B470, RegNetX-4.0GF71, and SE-ResNet-5072. The bar chart displays the count of topological dataset features for different classes within each model. The line chart indicates the count of common global features across different models, with each point on the horizontal axis representing common features for all models on the left. Error bars in both charts indicate standard deviation, capturing the variability across different samples within the dataset.
