The paper "Can Biases in ImageNet Models Explain Generalization?" by Paul Gavrikov and Janis Keuper investigates the relationship between biases in neural network models and their generalization performance. The authors focus on three specific biases: shape bias, spectral biases (low and high-frequency), and critical band bias. These biases are identified as factors that differentiate models from human vision and have been linked to various issues such as adversarial attacks and performance drops on distorted images. The study uses a large-scale analysis of 48 ImageNet models trained using different methods, all with the same ResNet-50 architecture, to understand how these biases interact with generalization. The generalization is assessed through multiple benchmarks, including in-distribution data, robustness to corruptions, conceptual changes, and adversarial attacks. Key findings include: - No single bias can fully explain generalization. - Shape bias is negatively correlated with in-distribution accuracy. - Spectral biases show mixed correlations, with high-frequency bias positively correlated with most aspects of generalization except for adversarial robustness. - Critical band parameters (bandwidth, center frequency, and peak-noise sensitivity) show complex relationships, with bandwidth positively correlated with in-distribution and robustness but negatively with adversarial robustness. The authors conclude that while some biases may be necessary for generalization, they alone are insufficient to explain it comprehensively. They also highlight the importance of rigorous experimental methodologies and caution against overreliance on bias measures without comprehensive validation.The paper "Can Biases in ImageNet Models Explain Generalization?" by Paul Gavrikov and Janis Keuper investigates the relationship between biases in neural network models and their generalization performance. The authors focus on three specific biases: shape bias, spectral biases (low and high-frequency), and critical band bias. These biases are identified as factors that differentiate models from human vision and have been linked to various issues such as adversarial attacks and performance drops on distorted images. The study uses a large-scale analysis of 48 ImageNet models trained using different methods, all with the same ResNet-50 architecture, to understand how these biases interact with generalization. The generalization is assessed through multiple benchmarks, including in-distribution data, robustness to corruptions, conceptual changes, and adversarial attacks. Key findings include: - No single bias can fully explain generalization. - Shape bias is negatively correlated with in-distribution accuracy. - Spectral biases show mixed correlations, with high-frequency bias positively correlated with most aspects of generalization except for adversarial robustness. - Critical band parameters (bandwidth, center frequency, and peak-noise sensitivity) show complex relationships, with bandwidth positively correlated with in-distribution and robustness but negatively with adversarial robustness. The authors conclude that while some biases may be necessary for generalization, they alone are insufficient to explain it comprehensively. They also highlight the importance of rigorous experimental methodologies and caution against overreliance on bias measures without comprehensive validation.