@Ryan, The “optimal” in optimal inference does not refer to a human’s absolute ability to perform in this task, only to how well he/she can incorporate the observations he/she has received. Imagine I ask you to play a game of Plinko. Except, because I am a neuroscientist, I have twisted the goal of the game to make it a bit more interesting. Rather than ask you to drop disks into the Plinko board, I drop them all from one particular slot at the top before you arrive. I call you in, and your job is to determine which location I dropped them from using their ending locations at the bottom.
Now, due to the inherent randomness in how the disks fall, you will never be able to tell exactly where I dropped them from. You will be able to make a good guess, however, because Plinko disks tend to end up more or less in a binomial distribution centered around the location I dropped them from. If you are aware of this distribution (or I train you to learn it), there is a unique, mathematically well defined way to determine the most likely location from which I dropped the disks.
Due to the noise in how the discs land, you will never be perfect in this task. There is, however, a best way to do this task, and that is the solution I told you about. If you know how the disk locations are determined (the binomial distribution), the best you can do is to use optimal inference. This is what the authors think the brain is doing: it represents the distribution from which the objects you see are drawn (a binomial distribution in this example), combines this distribution with the observations (disk locations), and makes the best possible guess it can.
Now to specifically answer your questions:
(1) Imagine I asked you to play this game, but rather than letting you see exactly where the disks are, I allowed you to look at it through foggy air and for only a fraction of a second. Now, not only is there randomness in how the disks actually land, but due to the external noise, there is also noise in where you think the disks are. I would not expect you (or any other creature) to do as well in this task as with good viewing conditions. Despite this handicap, if you know how the fog affects your vision (how it affects the distribution of light hitting your eyes), there is still a best, or optimal, way of doing the task. As long as you can know the distributions from which your observations come from, this method well defined.
(2) It is true that if you could simulate the exact physics of how the disks bounced around (and I don’t think anyone can), you could get a nearly perfect estimate, but there will always be some noise in your simulation and you will have to make a estimation at some point. With a bigger brain you could possibly represent the distribution of final disk locations better than with a binomial distribution. The experiment mentioned this article, however, controls very closely the distributions from which the images on the screen are drawn, so that the optimal strategy can be exactly determined. Having a bigger brain would not necessarily help in the task, since all it takes to do the optimal computation is knowing the distributions of the images and distribution of noise in the observations. The experiment does indeed find that humans, regardless of brain size or visual ability, are able to do these computations (or something very close to them) in order to make an optimal response.
For another explanation of optimal inference, see this tutorial.
@Paul, There are theoretical as well as practical implications of human optimality in perception. One goal of neuroscience is to understand how the brain is able to perform the complicated computations required to drive day-to-day behavior. If we observe a creature to behave optimally, we have a good clue that it employs certain specific computations with its brain, and we can peer inside the skull with neurophysiological techniques to see if this is in fact the case. Also, if we observe optimality in vision, there is good motivation to test other senses for the same behavior. Practically, this research has several applications. Neurology and psychiatry, for example, would benefit from a better understanding of brain function. It could be that strokes or lesions in specific parts of the brain cause deficits in such optimal computation, and therefore doctors could predict the effects of certain injuries or diagnose problems by testing the deficits in optimal behavior. Several psychiatric disorders have also been shown to cause specific, observable effects on visual perception, and beg for a better explanation of mechanism.
Artificial vision also stands to benefit from this research. Understanding how humans see the world will help the design of image recognition software. Humans have the amazing ability to understand the content of complicated, noisy pictures viewed very briefly, a task in which computer algorithms still require heavy development. It would be easy for you or I to look at a picture and identify what parts of the picture belong to what object. Even the most complex computer models, however, have large problems when dealing with mildly complicated images (imagine a dog seen through the gaps in a fence). With a better model one could write software to, for example, automatically tag images online with their contents or more accurately detect tumors in CT scans.
Source : http://blogs.discovermagazine.com/80beats/2011/05/10/humans-are-lean-mean-seeing-machines/Thank you for visit my website