Shared Visual Abstractions
This work suggests shared visual representations between human and computer vision systems, which could inform interpretability and cross-system generalization in AI.
The paper demonstrates that neural networks create abstract art forms that are broadly recognizable across different computer vision systems and by humans, with these drawings sometimes eliciting stronger label responses than validation set examples.
This paper presents abstract art created by neural networks and broadly recognizable across various computer vision systems. The existence of abstract forms that trigger specific labels independent of neural architecture or training set suggests convolutional neural networks build shared visual representations for the categories they understand. Computer vision classifiers encountering these drawings often respond with strong responses for specific labels - in extreme cases stronger than all examples from the validation set. By surveying human subjects we confirm that these abstract artworks are also broadly recognizable by people, suggesting visual representations triggered by these drawings are shared across human and computer vision systems.