Jeonghwan Cheon¹, Se-Bum Paik¹; ¹Korea Advanced Institute of Science and Technology

Infants have limited sensory capacities, including low visual acuity and color sensitivity. These limitations are often considered inevitable due to early sensory and cortical immaturity, though the functional role of this process remains unclear. In contrast, machine learning models typically maintain high sensory acuity throughout the learning process, often becoming trapped in local optima biased by peripheral properties in the training data. As a result, these models fail to generalize to unbiased data, where peripheral bias attributes have been removed. In this study, we demonstrate that initial sensory degradation can effectively guide neural networks to learn robust representations without being biased toward spuriously correlated peripheral attributes. Inspired by the developmental process of sensory systems, we trained neural networks on an initially grayscaled and blurred dataset, gradually increasing color sensitivity and spatial resolution during training, as proposed previously (Vogelsang et al., 2024). Specifically, we used spuriously correlated image benchmarks for training, such as Colored MNIST and Corrupted CIFAR-10, where biased attributes (i.e., color or corruption type) are strongly correlated with intrinsic attributes (i.e., shape or object). Both the conventional and initial degradation training schemes achieved comparable performance on bias-aligned samples. However, the initial degradation approach significantly outperformed the conventional scheme on bias-conflicting samples, where spuriously correlated peripheral biases are removed. We found that neural representations learned through conventional training often capture biased attributes, while those learned with initial degradation focus more on intrinsic attributes. Our findings suggest that initial sensory degradation is a key biological strategy that enables networks to learn debiased, robust neural representations without overfitting to biased attributes. This highlights a simple yet powerful biological strategy for learning robust neural representations and offers a potential solution to debiasing challenges in machine learning.

Acknowledgements: This work was supported by the National Research Foundation of Korea (NRF-2022R1A2C3008991 to S.P.) and by the Singularity Professor Research Project of KAIST (to S.P.).