Visualizing the Invisible: Generative Visual Grounding Empowers Universal EEG Understanding in MLLMs
Researchers have developed a new framework called Generative Visual Grounding (GVG) to improve the understanding of electroencephalogram (EEG) data using multimodal large language models (MLLMs). GVG addresses the scarcity of visually-evoked EEG datasets by generating proxy images from non-visual EEG signals. These generated images provide visual context, allowing MLLMs to leverage their visual priors for more effective interpretation of clinical states. Experiments show that this approach, even with a lightweight model, can match or exceed the performance of larger text-aligned models. AI
IMPACT This research could enable more accurate interpretation of brain activity data by leveraging visual priors in AI models.