Machines Like Us

Don't worry, the thought police are not coming for you. It's interesting, but what they have achieved is like guessing what's happening in a ballgame by listening to the crowd noise outside the stadium. That's not a perfect analogy, but the point is that you have to know a lot about the game and why the crowd sounds like it does in response to various events before you'd be able to make guesses about what's going on. Likewise, this imaging scenario looks at large-scale patterns of brain activity (akin to crowd noise), in which a pattern of brain activation becomes associated with a particular visual image (akin to learning why certain crowd noises correspond with certain game events). From the article:

The first step is to train the software decoder by scanning a subject's visual cortex while they view thousands of images over five hours.

A very key point here is that you have to train the thing for each new subject. We don't all have identical patterns of activation, because we all learn and acquire in idiosyncratic ways; each person's pattern of visual memory is unique, though there may well be certain commonalities.

One final thing. This is not, as some quoted in the article suggest, an advance in terms of understanding how the brain works. It is an advance, but only in our technological abilities. It's an advance of brain imaging, but more importantly, an advance in computer science, in that researchers have created software that can be trained to associate brain images (which is a lot of data) with actual black and white images (also a lot of data). That isn't to say that we know anything we didn't know before about how the brain encodes that information, just that we're better able to make inferences about what a particular brain state, as imaged by fMRI, suggests.