learning, unsupervised

This post is about image processing in the brain.

If you look at a digital image, the input is just a bunch of numbers (the red, green, and blue values for each pixel). The same is (sort of) true for the data your eyes collect from the world.

But, how does your brain go from this long list of numbers to the more abstract (and useful) representation "I am looking at my desk, with a laptop and a cup of coffee on it" (or whatever you happen to be looking at)?

There's a lot of stuff going on here that is just not yet known. This is also (incidentally), more-or-less what my PhD research is about.

What a lot of people (myself included) suspect is that the first few stages of image processing in the brain are just there to find common patterns, in a way that reduces redundancy. As an analogy, consider this line of text: thisisabunchofwordswithnospacesbutyoucanstillfigureitout

When your brain sees this, it "knows" what the common features are (words), and it picks them out of the slop. Then, the next stages of image processing (that do the abstractions, etc.) get these nice neat "words" to process instead of the (more complicated) raw input.

This process of finding the common patterns in a bunch of data is called unsupervised learning because there's no "teacher" signal saying "look for the red blob" (or whatever): you really just look around and find patterns that occur the most often.

If the early visual system does this sort of thing, then people should be able to write computer programs to find the common patterns in natural scenes, and use those to predict some of the properties of the visual center(s) of the brain. Indeed, several of the guys in our theory center built their careers on doing just that, with great success.

These same techniques are useful in other fields that seek to find patterns in data, like finance (looking for stocks that are likely to behave similarly, for example).