NEvo is a research project that searches for AI-generated videos that maximally activate a chosen visual brain region. The EPFL and Johns Hopkins project page describes it as Neural-Guided Evolutionary Video Synthesis. It starts by training an encoding model, effectively a digital twin that predicts how visual areas respond to video, and then uses that prediction as a reward signal.

The search process is close to an evolutionary loop. A candidate video is represented through genes such as subject, lighting, motion, and mood. NEvo generates a batch, scores each candidate with the digital twin, keeps the strongest ones, then mutates and recombines them across generations. To keep the search manageable, it first finds a strong still image and then runs a second search over motion to turn it into a two-second clip.

As a scientific instrument, the appeal is clear. Instead of hand-picking stimuli from a researcher’s prior assumptions, NEvo searches a richer space and checks what a model of the visual brain predicts. The project page says synthesized clips line up with known selectivity across regions, including faces, places, bodies, motion, patterns, and social scenes. That makes it a way to probe visual selectivity with less experimenter bias.

The HN discussion centered on the same mechanism from another angle. Some commenters immediately saw the path from research tool to optimized visual superstimulus, especially in a world where recommendation systems already learn which short videos hold attention. Others pushed back that the paper should be read as a neuroscience tool, not a consumer engagement product. Both reactions are useful because they identify the real hinge: once brain response becomes an optimization target, the boundary between measurement and manipulation needs to be explicit.

HN discussion · Source

#neuroai

NEvo Searches for Videos That Maximally Activate a Brain Region