WOW! Scientists have developed a system to capture visual activity in human brains and reconstruct it as digital video clips

From:
Gizmodo

The left clip is a segment of the movie that the subject viewed while in the magnet. The right clip shows the reconstruction of this movie from brain activity measured using fMRI. The reconstruction was obtained using only each subject's brain activity and a library of 18 million seconds of random YouTube video. (In brief, the algorithm processes each of the 18 million clips through the brain model, and identifies the clips that would have produced brain activity as similar to the measured brain activity as possible. The clips used to fit the model, those used to test the model and those used to reconstruct the stimulus were entirely separate.) Brain activity was sampled every one second, and each one-second section of the viewed movie was reconstructed separately.

UC Berkeley scientists have developed a system to capture visual activity in human brains and reconstruct it as digital video clips. Eventually, this process will allow you to record and reconstruct your own dreams on a computer screen.

I just can't believe this is happening for real, but according to Professor Jack Gallant—UC Berkeley neuroscientist and coauthor of the research published today in the journal Current Biology—"this is a major leap toward reconstructing internal imagery. We are opening a window into the movies in our minds."

Indeed, it's mindblowing. I'm simultaneously excited and terrified. This is how it works:

They used three different subjects for the experiments—incidentally, they were part of the research team because it requires being inside a functional Magnetic Resonance Imaging system for hours at a time. The subjects were exposed to two different groups of Hollywood movie trailers as the fMRI system recorded the brain's blood flow through their brains' visual cortex.

The readings were fed into a computer program in which they were divided into three-dimensional pixels units called voxels (volumetric pixels). This process effectively decodes the brain signals generated by moving pictures, connecting the shape and motion information from the movies to specific brain actions. As the sessions progressed, the computer learned more and more about how the visual activity presented on the screen corresponded to the brain activity.

An 18-million-second picture palette

After recording this information, another group of clips was used to reconstruct the videos shown to the subjects. The computer analyzed 18 million seconds of random YouTube video, building a database of potential brain activity for each clip. From all these videos, the software picked the one hundred clips that caused a brain activity more similar to the ones the subject watched, combining them into one final movie. Although the resulting fukn is low resolution and blurry, it clearly matched the actual clips watched by the subjects.

Think about those 18 million seconds of random videos as a painter's color palette. A painter sees a red rose in real life and tries to reproduce the color using the different kinds of reds available in his palette, combining them to match what he's seeing. The software is the painter and the 18 million seconds of random video is its color palette. It analyzes how the brain reacts to certain stimuli, compares it to the brain reactions to the 18-million-second palette, and picks what more closely matches those brain reactions. Then it combines the clips into a new one that matches what the subject was seeing. Notice that the 18 million seconds of motion video are not what the subject is seeing. They are random bits used just to compose the brain image.

Given a big enough database of video material and enough computing power, the system would be able to match any images in your brain.

In this other video you can see how this process worked in the three experimental targets. On the top left square you can see the movie the subjects were watching while they were in the fMRI machine. Right below you can see the movie "extracted" from their brain activity. It shows that this technique gives consistent results independent of what's being watched—or who's watching. The three lines of clips next to the left column show the random movies that the computer program used to reconstruct the visual information.

Right now, the resulting quality is not good, but the potential is enormous. Lead research author—and one of the lab test bunnies—Shinji Nishimoto thinks this is the first step to tap directly into what our brain sees and imagines:

Our natural visual experience is like watching a movie. In order for this technology to have wide applicability, we must understand how the brain processes these dynamic visual experiences.

The brain recorders of the future

Imagine that. Capturing your visual memories, your dreams, the wild ramblings of your imagination into a video that you and others can watch with your own eyes.

This is the first time in history that we have been able to decode brain activity and reconstruct motion pictures in a computer screen. The path that this research opens boggles the mind. It reminds me of Brainstorm, the cult movie in which a group of scientists lead by Christopher Walken develops a machine capable of recording the five senses of a human being and then play them back into the brain itself.

This new development brings us closer to that goal which, I have no doubt, will happen at one point. Given the exponential increase in computing power and our understanding of human biology, I think this will arrive sooner than most mortals expect. Perhaps one day you would be able to go to sleep with a flexible band around your skull labeled Sony Dreamcam, wirelessly connected to your iPad 7. [UC Berkeley]

This video is organized as follows: the movie that each subject viewed while in the magnet is shown at upper left. Reconstructions for three subjects are shown in the three rows at bottom. All these reconstructions were obtained using only each subject's brain activity and a library of 18 million seconds of random YouTube video that did not include the movies used as stimuli. (In brief, the algorithm processes each of the 18 million clips through the brain model, and identifies the clips that would have produced brain activity as similar to the measured activity as possible. The clips used to fit the model, the clips used to test the model and the clips used to reconstruct the stimulus were entirely separate.) The reconstruction at far left is the Average High Posterior (AHP). The reconstruction in the second column is the Maximum a Posteriori (MAP). The other columns represent less likely reconstructions. The AHP is obtained by simply averaging over the 100 most likely movies in the reconstruction library. These reconstructions show that the process is very consistent, though the quality of the reconstructions does depend somewhat on the quality of brain activity data recorded from each subject.

You can find more information about this work at our laboratory web site: http://gallantlab.org