Content area
The field of psychology has long recognized a “basic level” of categorization that humans use when labeling visual stimuli, a term coined by Rosch in 1976. This level of categorization has been found to be used most frequently, to have higher information density, and to aid in visual language tasks with priming in humans. Here, we investigate basic-level categorization in two recently released, open-source vision-language models (VLMs). This thesis demonstrates that both Llama 3.2 Vision Instruct (11B) and Molmo 7B-D exhibit a preference for basic-level categorization, aligning with human behavior. Moreover, the models’ preferences are consistent with nuanced human behaviors like the biological versus non-biological basic level effects and the well-established expert basic level shift. The model also shows a increased performance on visual question answering tasks that is associated with its use of the basic-level. However, the benefit of expertise that is typically associated with a shift away from the basic level was not observed in our expert-prompting analysis. This thesis will contribute by demonstrating that:
1. VLMs tend to prefer categorization at the basic-level.
2. VLMs show a notable distinction in basic-level categorization frequency between biological and non-biological objects, consistent with known human-categorization behavior.
3. VLMs exhibit a similar decrease in basic-level usage frequency when prompted with expertise to that shown in human experts.
4. VLMs demonstrate basic-level usage and improved accuracy when prompted with the basic prompt.
5. VLMs do not present increased accuracy from expert prompting when they diverge from the basic-level.