AI Summary: Discovers that multimodal models like CLIP naturally develop 'multimodal neurons' that abstractly link images, text, and sketches of a concept, revealing both profound cognitive similarities to the human brain and novel adversarial vulnerabilities.
We investigate the internal representations of the CLIP model and discover the presence of 'multimodal neurons'. These neurons fire not only for specific visual features (like a spider) but also for the text representing that concept, and even abstract sketches or comic depictions of it. We demonstrate that these networks naturally develop representations akin to the 'Halle Berry neuron' discovered in the human brain, grouping disparate visual and textual stimuli under single, highly abstract conceptual nodes. Furthermore, we reveal how these abstract representations make the model vulnerable to a novel typographic attack, where simply writing a word on an object can completely alter the model's classification.
Share your opinion to help other learners triage faster.
Write a reviewInvite someone by email to share an invited review for Multimodal Neurons in Artificial Neural Networks.