Research Explainer · XiaoHu Explains

What does each part of your brain love to watch? EPFL made AI generate videos over and over until each brain region gave up its favorite clip

All results are computer-simulated predictions from a "digital twin" brain model — not yet validated with real human brain imaging.
Quick Take
  • EPFL (École Polytechnique Fédérale de Lausanne), together with Johns Hopkins University, released NEvo: a system that uses an evolutionary algorithm to generate AI videos specifically designed to maximize activation in a target region of the visual cortex.
  • The system first trains a "digital twin" encoding model that predicts the brain's visual responses (backbone: Meta's V-JEPA 2), then runs a genetic algorithm to search the prompt space of a video generation model for the clip that drives the highest predicted activation in the target brain region.
  • The synthesized videos' predicted activation lands in the top 0.2% of responses to real videos from the Moments in Time library (beating 99.8% of real videos), reaching 95.8% of the response evoked by professionally designed localizer stimuli.
  • A "searchlight" scan along the brain's lateral pathway — from V1 to the anterior superior temporal sulcus (aSTS) — reveals a continuous gradient: preference for texture and color gradually shifts into a preference for motion, then into a preference for social interaction.
  • The paper is explicit that all these activation scores are computer-simulated predictions from the digital-twin model itself, not measurements from real human brain imaging — the authors call for future closed-loop experiments in real humans to validate them.
1Background

AI-synthesized videos, built to light up the brain

A research team from EPFL (École Polytechnique Fédérale de Lausanne) and Johns Hopkins University recently published a system called NEvo: it uses an evolutionary algorithm to generate AI videos, using them to map out the preferences of the brain's visual regions.

This is the first framework to use an AI encoding model to guide a systematic study of brain selectivity under natural, dynamic visual stimuli. Instead of hand-picking images, researchers let AI evolve videos that precisely light up a specific brain region.
◆
How strong are the synthesized videos? Predicted activation lands in the top 0.2% of responses to real videos (beating 99.8% of real videos), reaching 95.8% of the response evoked by professionally designed localizer stimuli. In other words, the AI-evolved stimuli match — and even exceed — the strongest responses among a large pool of real videos and expert-designed stimuli.

The accompanying paper, "NEvo: Neural-Guided Evolutionary Video Synthesis for Dynamic Visual Selectivity" (arXiv:2607.02317), comes with a project page where you can hover over each brain region to view the video synthesized specifically for it.

Six authors (Yingtian Tang, Sogand Salehi, Amir Zamir, and Martin Schrimpf are all at EPFL; Ming Zhou and Leyla Isik are at Johns Hopkins). Martin Schrimpf is the author of Brain-Score, a benchmark for evaluating brain-model alignment; Leyla Isik researches social visual perception.
2Neuroscience Background

How the brain sees a moving world: a gap still unfilled

The brain's visual system has a well-defined division of labor — classically described as two pathways: the ventral stream for recognizing objects, and the dorsal stream for handling motion and action. But more recent research has found a third: one dedicated to processing social information, and currently the least understood.

01
Ventral Streamventral stream

Recognizes objects, faces, scenes — answers "what is this."

02
Dorsal Streamdorsal stream

Handles motion and guides action — answers "where" and "how it moves."

03
Lateral Streamlateral stream

Processes biological motion, others' actions, social interaction. A newly identified pathway, the least studied of the three — exactly what NEvo is going after.