Modeling Human Concepts with Subspaces in Deep Vision Models

Improving the modeling of human representations of everyday semantic categories, such as animals or food, can lead to better alignment between AI systems and humans. Humans are thought to represent such categories using dimensions that capture relevant variance, in this way defining the relationship between category members. In AI systems, the representational space for a category is defined by the distances between its members. Importantly, in this context, the same features are used for distance computations across all categories. In two experiments, we show that pruning a model's feature space to better align with human representations of a category selects for different model features and different subspaces for different categories. In addition, we provide a proof of concept demonstrating the relevance of these findings for evaluating the quality of images generated by AI systems.