快猫短视频

BLEE-blee-bleeping saturates my earplugs. I cannot speak or shift my head a millimetre. I am not even supposed to move my eyes.

Lying here in a narrow tube, I am isolated except for a mirror tilted above my face. Reflected in it I see a computer screen with a grid of criss-crossing lines. That screen is my link to the researchers in the next room, who are gawking at pixelated slices of my brain as a functional magnetic resonance imaging (fMRI) scanner yanks them out of my head, 14 per second.

I work hard for those researchers, block out the noise from the machine, fasten my eyeballs to those lines and hold steady as tears well up. Focusing my attention on one set of lines at a time 鈥� left-leaning lines, right-leaning lines, then back to the left-leaning ones 鈥� I remind myself to blink and breathe.

It sounds like the movie The City of Lost Children, in which a mad scientist harvests the dreams of captive kids 鈥� except I鈥檓 here by choice. The researchers are watching patterns of blood flow in my brain, pinpointing spots where neurons are most vigorously stuttering out electric pulses. By reading subtle changes in my brain activity, these guys actually think they can eavesdrop on my thoughts. They think they can tell what I鈥檓 looking at, and which lines I鈥檓 paying attention to.

鈥淲ith fMRI, we can read out what a person is seeing or perceiving,鈥� says Frank Tong, the cognitive neuroscientist running the experiment here at Vanderbilt University in Nashville, Tennessee. 鈥淚t demonstrates that very basic forms of mind reading can be achieved with brain imaging.鈥�

Today鈥檚 exercise is just a proof of concept, but it opens up a world of possibilities. Sometime in the future, people who cannot move or speak might communicate using brainwaves picked up by a scanner. Perhaps you鈥檒l awaken and project your dreams onto a TV while sipping cappuccino. Or download from your brain, directly to DVD, the images of yaks and yurts that you saw during a holiday trek in Mongolia.

Until recently, mind reading was the preserve of quacks, but neuroscience has made respectable inroads. Researchers have tested brain-machine interfaces that allow people to control a cursor on a computer screen using scalp electrodes or brain implants. Studies in monkeys, and in people undergoing neurosurgery, show that if you slip electrodes into the brain, you might 鈥� depending on the neurons you hit 鈥� be able to determine if someone is looking at a picture of a car, the Eiffel Tower or the actress Jennifer Aniston (see 鈥淥pening the box鈥�). Scalp electrodes are not always reliable, however, and no one wants a hole in their head. This fMRI stuff, on the other hand 鈥� if it works on me, it might just work on you.

Tong hopes to use mind decoding, as he calls it, to explore the basis of visual consciousness. How do we perceive images in our heads, whether real or imagined? The process arises from the exchange of nerve pulses 鈥� a million-way conference call among neurons that changes every second: as your gaze shifts from this article to, say, the dog nuzzling your leg, some neurons drop off and others join in. Tong and others are trying to identify which brain areas participate and exactly what information they contribute.

As they prepared the experiment this morning, I felt like an astronaut 鈥� a neuronaut, you might say 鈥� getting ready for launch. Because the MRI magnet can turn loose metal objects into lethal projectiles, I was carefully de-metalled beforehand. Tong and three colleagues mounted a cage-like antenna over my head, standard for capturing MRI signals, and placed a custom-fit bite bar in my mouth. Turning screws, they fixed the bar and my head firmly in place.

An hour later the experiment is under way, and I could easily panic, immobilised like this, except I鈥檓 thoroughly engrossed in the task. The lines I鈥檓 looking at are changing thickness every few seconds. I report what I see by pressing buttons with my right hand. Thin, press 1. Thick, press 2. That鈥檚 how the researchers verify which lines I鈥檓 focusing on. Just as I鈥檓 hitting the Zen zone and those lines are popping out of the screen like neon licorice sticks, they disappear, and a text box appears 鈥� a message from Tong:

鈥淕ood job on that run. We see your brain. You have a very straight calcarine sulcus.鈥�

They tease their subjects like this, to prevent them dozing in the tube. My calcarine sulcus, by the way, is a fold in the rear of my brain, and it is central to this experiment.

That fold is part of the cerebral cortex, the furrowed grey matter that can be flattened out to display a map of our sensory world. Every square centimetre of skin, every piece of visual field and every odour has a speck of cortex devoted to it. The map would cover a large chessboard and is crumpled like origami to fit inside your skull. The calcarine sulcus contains much of the visual cortex, which receives signals from the retina of the eye via the optic nerve and lateral geniculate nucleus. The visual cortex identifies things like edges and line orientations, which go on to other pieces of cortex to be assembled into visions of broccoli, Hugh Grant, and so on.

Tong will analyse the activity in my visual cortex and try to decode whether I鈥檓 focusing my attention on the left or right-leaning lines. To explain how it works before we start, he pulls up an image on his computer: a quilt pattern as random as splotches on a technicolour cow. It is an enlarged view of the visual cortex, which contains thousands of 鈥渙rientation columns鈥�, each composed of thousands of neurons that respond sharply to lines oriented at a particular angle.

鈥淭hey can tell what I鈥檓 looking at by using fMRI to image my visual cortex鈥�

It should be possible for Tong to tell which lines I鈥檓 attending to, left-tilted or right-tilted, by imaging the cortex to see which columns of neurons light up 鈥� the ones that correspond to 45 degrees left or 45 degrees right. The problem is resolution: each column is just half a millimetre across, while the smallest chunk of brain that can be analysed reliably by fMRI is 3 millimetres on a side. On the scanner, these chunks are known as volume pixels, or voxels, and each one spans 30 or so orientation columns.

In 2004, while Tong was at Princeton University, he and his colleague Yukiyasu Kamitani found a way round this. Even if a voxel covers 30 columns, they reasoned, it probably contains slightly more of one type than others, so overall it will respond slightly more to one line orientation than to others. If you had to guess which of two lines I was looking at, based on one voxel, you might be right only 52 per cent of the time. 鈥淚t鈥檚 a tiny, tiny bias,鈥� says Tong. 鈥淏ut if I pulled together enough of these guys, I might get up to 90 per cent on my predictions.鈥�

How does he pull them together?

鈥淲e typically grab 400 to 800 of these voxels,鈥� he says, which encompasses most of the visual cortex. In the first half of the experiment, they show me lines oriented at known angles and determine the bias of each voxel (see Diagram). In the second half, they use that information to decode which lines I am looking at, using an algorithm that polls the response of all voxels and weights each one by the confidence of its response in the first phase. A voxel with 60 per cent bias during calibration carries more weight than one with 52 per cent bias. Adding up the voxels gives a prediction that is the best statistical match.

After two hours in the cramped tube, I am spent. I stagger zombie-limbed from the machine. As Tong and I walk back to his office, I ask him if he thinks the experiment worked. It depends, he says. Even very minor eye movement or inattention can muddle the signal, which is why he normally uses experienced subjects. It will require five hours of data crunching before he knows whether he could read my mind.

In published results, Tong and Kamitani were able to predict correctly 56 per cent of the time which of eight orientations of lines people were seeing, compared with 12.5 per cent for chance. When subjects were shown a grid of criss-crossing lines, as I was, the researchers predicted correctly 80 per cent of the time which lines were being attended to (Nature Neuroscience, vol 8, p 679).

When they first presented the findings at the 2004 Vision Sciences Society meeting in Sarasota, Florida, it created shock waves. 鈥淧eople would have thought they were insane for trying it,鈥� says neuroscientist John-Dylan Haynes of the Max Planck Institute for Human Cognitive and Brain Sciences in Leipzig, Germany. 鈥淓verybody thought the resolution of fMRI was far too low.鈥�

Within a few months, however, Haynes and Geraint Rees, both then at University College London, had successfully used mind decoding to read orientations of lines that subjects were not even conscious of (Nature Neuroscience, vol 8, p 686). 鈥淚f you鈥檙e really interested in squeezing every bit of information from someone鈥檚 brain,鈥� says Haynes, 鈥渢hen you鈥檙e going to use these techniques.鈥�

Tong is now pushing the technique to decode not just lines, but everyday things. He is analysing fMRI signals from the inferior temporal cortex, a part of the brain that identifies objects. One experiment involves showing people different kinds of birds. 鈥淲e can tell penguins and pigeons apart,鈥� he says. However, it is not clear exactly what that means. The scanner spots activity patterns that correlate with pigeon or penguin, but cannot shed any light on why the brain responds that way. 鈥淲e don鈥檛 know if it鈥檚 鈥榖irdness鈥� we are telling apart, or just general shapes. It could be something really deep that we鈥檙e decoding, or it could be a trivial difference,鈥� Tong says.

Nevertheless, the ability to tell objects apart seems promising. What about more complex scenes? If I saw an elephant walking past the Taj Mahal, could Tong decode it? 鈥淲e would have to look at all of the brain activity we recorded whenever you thought about elephants,鈥� he says, 鈥渁nd when you saw something walking, and when you saw the Taj Mahal. It would be pretty darn hard.鈥�

Haynes agrees. 鈥淲hat we want to do is take someone from the street, put them in the brain scanner and say, you are thinking this,鈥� he says. 鈥淏ut you have to train the algorithm for every type of thought you could possibly have, and that鈥檚 not practical.鈥� That鈥檚 the trouble: you can鈥檛 expect people to lie inside an fMRI machine 24/7. 鈥淚f in the future we get some very handy portable imaging device, we can record brain activity in daily life,鈥� says Kamitani, who now runs his own lab at ATR Computational Neuroscience Laboratories near Kyoto, Japan.

What鈥檚 more, fMRI has fundamental limitations. A few millimetres on a side per voxel is the best resolution available, but that still lumps together about a million neurons. The resolution in time is also bad: while electrodes planted directly in the brain can register millisecond nerve pulses, the blood flow changes measured by fMRI take several seconds to materialise.

Compare all this, says Haynes, to a scene from the 1976 movie Futureworld, in which a woman鈥檚 dreams are projected on TV in real time. 鈥淭hat鈥檚 the kind of fantasy everyone in this field has,鈥� he says. 鈥淏ut if that鈥檚 the device you want to build, then I鈥檓 pretty sure you need to record from single neurons.鈥� You might also need to record from several brain regions at once, including the auditory cortex (Science, vol 309, p 951). While researchers have managed to record simultaneously from a few hundred electrodes, no one knows how to record from millions of single neurons, let alone do it non-invasively.

If it could be done, though, what would be our chances of decoding dreams and thoughts? Such processes should activate the visual centres of the brain, but the signals are 鈥渘oisier and weaker鈥�, Tong says. 鈥淚t presents a greater challenge, but it should be doable in many cases.鈥� Kamitani concurs. 鈥淚t鈥檚 possible,鈥� he says. 鈥淲e have tried people just thinking about lines, and in some people you can decode above chance level. It seems to depend on how well you can imagine a visual image.鈥�

鈥淚t should be possible to decode a person鈥檚 dreams and thoughts, but the signals are noisier and weaker鈥�

A mind鈥檚-eye decoder might require a lot of training to use, but one could imagine building a communication device that could broadcast the thoughts of paralysed people on a computer screen. On another front, people with schizophrenia might one day wear brain monitors that notify their doctors of hallucinations and even allow others to download them for clinical study.

A growing concern is what else might be done with all this neural information (快猫短视频, 31 July 2004, p 38). 鈥淭he technologies are evolving into hands-off techniques,鈥� says Judy Illes, a neuroethicist at Stanford University. 鈥淲e鈥檒l be able to get brain measurements without touching somebody, and even without their consent.鈥� Could identity thieves use brain scans to steal bank account numbers? Will governments conduct neural interrogations? Though these prospects seem remote, we need to ensure that mind decoding is developed in a way that respects privacy, says Illes.

For now, rest assured. Researchers are a far cry from being able to read out anything as specific as an account number, and they have to calibrate the software for every person. It鈥檚 one of the mysteries of visual consciousness: the same information seems to be represented differently in every brain. 鈥淚t鈥檚 not like we can scan your brain and then plug in another person鈥檚 brain and decode it,鈥� Tong says.

A few days later, I call Tong to hear the results of my experiment. He is upbeat. 鈥淲hen we tried to classify which of the two gratings you were paying attention to,鈥� he says, 鈥渨e could do it with 90 per cent accuracy.鈥� In other words, they could tell exactly which lines I was looking at 9 times out of 10. 鈥淵ou were really on task,鈥� he says.

As I hang up, I鈥檓 strangely glad to know my brain is neuro-legible. Reading minds may be hard work, but having your mind read is even harder. Still, it is comforting to know that my mind鈥檚 eye could only be read with my help. It gives me hope that our neural privacy can be protected, leaving us free to enjoy the finer fruits of decoding.