Cinematic Mindscapes: High-Quality Video Reconstruction from Brain Activity

arxiv.org

1 pointftxbro1 comment

The recovered videos were evaluated with various semantic and pixel-level metrics. We achieved an average accuracy of 85% in semantic classification tasks and 0.19 in structural similarity index (SSIM), outperforming the previous state-of-the-art by 45%.

"It's a cat, moving" and "meh. Two people walking is good semantic match for person walking" levels of semantic match. (Both examples in the paper)

Structural similarity.. frankly it's bullshit. It's not even close for fidelity to subject by any human measure.

Still impressive but don't over dream it.

They should rerun the fMRI in 6/9/12/24 months with co founding imagery to test subjects octopus moving, six clowns walking. See how good retention of core semantics is if nothing else.

Reply on news.ycombinator.com