Weizmann研究所Brain-IT用fMRI扫描重建所见图像
An AI system called Brain-IT can reconstruct recognizable images from brain scans of people looking at pictures. Developed at the Weizmann Institute, it predicts both image content and layout, then uses a diffusion model to generate the reconstruction. Researchers expanded its training data with simulated brain activity. With one hour of scans from a new person, it achieved results comparable to earlier systems trained on 40 hours. But it still makes mistakes: one reconstruction turned a dog into a goat. Reconstructing dreams is a future ambition. For now, the system reconstructs viewed pictures, with impressive results that remain imperfect predictions. Study: “Brain-IT: Image Reconstruction from fMRI via Brain-Interaction Transformer” DOI: 10.48550/arXiv.2510.25976
Weizmann研究所开发出Brain-IT系统，可从人看图片时的fMRI脑扫描中重建出可辨认的图像。该系统分别预测图像的内容与布局，再用扩散模型生成重建结果，并用模拟脑活动扩充训练数据。
用视频里狗变山羊的失败案例和1小时扫描对比40小时的数字，讲清Brain-IT如何从fMRI重建所见图像。
粉丝 8.1万
The latest in AI and robotics, explained Founder & CEO of The Rundown AI Zurich 🇨🇭 / NYC 🇺🇸
- 浏览
- 7361
- 点赞
- 411
- 评论
- 13
- 分享
- 48
- 收藏
- 87
数字来自平台，为最近一次抓取时的数据。
以下是视频口播的自动转写，可能有错字。
This ai can turn a brain scan into a picture of what you're looking at. It can reconstruct both the objects you saw and how they were arranged.
Researchers developed an ai system called brain it that analyzes fmri scans which track changes in blood oxygenation associated with brain activity.
Basically, one branch predicts the image's layout and colors, and another predicts its content like a bunch of bananas on a plate.
And then those clues guide a diffusion model to generate a reconstruction. But there was a problem with this, because collecting brain scans is very expensive.
So the team trained another model to work backwards. It takes an ordinary image and predicts the brain activity someone might produce while looking at it.
And those simulated recordings provide extra training examples without putting anyone inside a scanner. And with just one hour of scans from one new person, the researchers reported results comparable to earlier systems trained on 40 hours.
But these images are still predictions. In one example, a dog in a bathtub became a goat. But it's very fascinating because the researchers hope to eventually reconstruct entire.
Entire dreams, but for now, they've still demonstrated something very remarkable, turning brain activity into recognizable versions of pictures people actually saw.
If you want to stay in the loop with the latest in ai and robotics, follow for more.
来源：TikTok · Rowan Cheung · tiktok.com