Abstract
How do we tell that one bird on a tree is a sparrow and the other is a warbler? Humans recognize visual objects by processing a hierarchy of low- to high-level visual information, but the involvement of low-level information in object categorization remains to be explored. Unlike higher-level information (e.g., shapes and textures), low-level information (e.g., spatial orientations and frequencies) diagnostic of object categories can be challenging to capture in naturalistic images. Here, we aimed to leverage category-diagnostic low-level visual information to optimize human category learning – particularly the learning of unfamiliar bird categories. Specifically, we used a variant of the Spatial Envelope model to represent naturalistic bird images as sets of low-level features based on oriented Gabor filters. Then, we trained a classifier to categorize these low-level bird representations. From the trained classifier, we obtained weights of the low-level features to create weighted masks that selectively added noise to the diagnostic or non-diagnostic low-level information in each image. Subsequently, we randomly assigned 48 participants to learn to categorize the bird images with masked diagnostic or masked non-diagnostic information. Compared to the masking of diagnostic information, the masking of non-diagnostic information resulted in a steeper learning slope and a greater speed-up in reaction time for correct learning trials. When participants categorized novel bird images after learning, the masking of non-diagnostic information led to faster responses in correct trials relative to the masking of diagnostic information. In conclusion, our findings revealed that low-level visual information defining specific categories can be extracted from naturalistic object images. Furthermore, we demonstrated that diagnostic low-level information can be leveraged to optimize learning of naturalistic object categories and support generalization to novel objects from those categories.