Brain algorithms

[mirhagk]

I've been very interested in how the brain does various tasks as of late. Is the brain a better algorithm, merely heuristic guesses for things, or is genuinely a super computer. I'd like to share and hear people's thoughts on this, and results for various computer vs human implementations of algorithms.

Here's an interesting one I found:http://web.uct.ac.za/depts/psychology/psy300/sternb.html It's called the Sternberg experiment, and basically it is checking:

code:

List<int> numbers=new List<int>(); //get numbers int num=//some random number if (numbers.Contains(num)) { }

The people are supposed to respond as quickly as they are 100% sure. The results of the experiment have many implications, but one in particular stands out to me. The amount of time it takes to respond is determined to be
RT = 397.2 + 37.9n
where RT is the response time and n is the number of items in the list the person was memorizing. Note also that the response time is unrelated to whether the item is actually in the list or not, which means every item is compared, there is no lazy evaluation.

This is very interesting results, because it appears that in the case of this specific algorithm, our brain takes the most naive approach to it, searching every single item, and then once done checking if it is in there or not. From what I can figure the formula would be like:

code:

bool found=false; foreach(int number in numbers)     if (number == numLookingFor)        found=true; return found;

There is no parallel programming here, and no heuristics, just plain, naive code. Note that it also takes 40ms to check against each number, which is pretty slow in comparison to computers. Mind you that memorization and data comparison is not one of the brain's highlights, but nonetheless you'd expect the algorithm to be a little bit better, even if the data was not always perfectly memorized.

If anyone else has any information on other algorithms, or would like to discuss how the brain might compute various things, please feel free to add, and if you have comments or suggestions to my theory here, or anyone else's, please comment, as this area is very intriguing to me (and useful to the fields of neuroscience and artificial intelligence)

[Insectoid]

All heuristics. Brains are fast but unreliable. Your brain fills in the blanks in hearing and vision a lot, but it isn't accurate, which is why we see & hear things that aren't there, or are different from reality (see optical/audial illusions). Of course, it's right (or close to right) often enough that we don't care.

It's for this reason that I don't think speech recognition will ever reach natural levels. How are we supposed to make a computer understand what we're saying when we can't even do it ourselves half the time? We can get close to it, but it will never, ever be perfect. I think more time should be spent on the responses than currently is. I'd get much less frustrated with my computer if it would get frustrated with me. But now I'm getting off-topic.

Brains are fast, but inaccurate.

[mirhagk]

But it's not always heuristics, as shown in my example. Also what might the heuristics be? i think emotion is a big one, but what about verbal understanding as you bring up, because heuristics are already used, but what are they incapable? I don't use them not because they suck but rather it's awkward to talk to your computer, which suggests a social problem, not a technical one.

[DemonWasp]

I'm curious whether the performance changes if the question is less abstract than number memorization (even if analogous).

Consider this problem: You are given a list of names of your friends that are at a party. You are asked "Is (name) at the party?"

Would the answers to that be any different?

[mirhagk]

hmm, well the said study has been done more abstractly for animals, although I'm not sure if something close to the example you mention has been done. I kinda want to see what the results would be like, because I imagine the brain being able to optimize that.

Then again I thought the average person would respond faster than 440 ms to a scenario with a single comparison, so intuition isn't always a good judge.

EDIT: I would also like to see this study down a little better, as the original experiment had a relatively low number of people involved, even though it had a high number of trials. Working for a clinical trial building has really opened my eyes to stuff like this in studies.

[mirhagk]

Sorry for the double post but is anyone here geniunely interested in doing some studies related to this? They could be done by creating a program for it, and then recording people's results. It might be interesting, and we could control all sorts of variables (for instance does counting down to the question using numbers affect the results etc). If anyone is interested let me know, as I think I'm gonna draw up some ideas for a program to measure this.

[Aange10]

I'd be very interested in the product. It seems like a doable thing, and I have a lot of friends who would be willing to take the tests just for the sake of making data.

[mirhagk]

Sweet, did you want to help in the design/construction of the program at all?

[Aange10]

Indeed; We can communicate on skype.

[trishume]

The general rule for humans vs. computers when performing different tasks is this:
Would this be trivial for a 6 year old?

If a 6 year old can do it without any effort then it is often very difficult for computers.
This is because the subconscious behaviors like sight and language are built into us by evolution whereas higher level logic is a learned skill.

Checking if a number is in a list is not something that evolution has tuned us to do well.
If you changed the task slightly from numbers to faces, humans would cream computers.

This will be obvious for many people but some people it is surprising.
People think that the same things that are hard for us are hard for the computer, like chess, when often it is the opposite.

[mirhagk]

Chess is actually very difficult for machines. It's a very hard problem that we have eventually solved with a computer, but humans are much better, other than the fact that computers work much harder. The original post is interesting because it shows that computers don't just have the advantage of a faster processor, they show that they have a better algorithm for this, and that memory matching from short term memory (at least for numbers) is done linearly, with no lazy evaluation, which means there is no multithreading. The key point isn't that a computer is faster, it's that a computer is more efficient, has a smarter algorithm for it.

Faces would be an interesting thing to test with this test, and I'll include it in my program to see if your hypothesis is true.

[trishume]

mirhagk @ Sun Aug 19, 2012 1:10 pm wrote:

Chess is actually very difficult for machines. It's a very hard problem that we have eventually solved with a computer, but humans are much better, other than the fact that computers work much harder. The original post is interesting because it shows that computers don't just have the advantage of a faster processor, they show that they have a better algorithm for this, and that memory matching from short term memory (at least for numbers) is done linearly, with no lazy evaluation, which means there is no multithreading. The key point isn't that a computer is faster, it's that a computer is more efficient, has a smarter algorithm for it.

Faces would be an interesting thing to test with this test, and I'll include it in my program to see if your hypothesis is true.

I know that chess is difficult computationally. However, compared to things like speech recognition, chess is child's play.

[mirhagk]

Despite countless years of research into building chess AI, we still haven't gotten it to an efficient system, and humans are still MUCH more efficient at chess then computers are. Speech recognition is actually becoming more advanced than chess right now, because speech recognition can be analyzed and compared very easily, and the system can "learn" a lot easier. On the other hand chess programs don't learn very well, they are either brute force machines that can prune options a little, or they are pre-loaded with game winning strategies developed by humans.

It is true that speech recognition is harder to get started in (there's no straightforward brute force algorithm that you can reasonably program), but to make something efficient chess is probably just as hard, if not harder than speech recognition.

[evildaddy911]

i believe that much of recognition (ie. sound) is heuristic. everyone's voice sounds different, so picking out words would be difficult to do AND keep up with the speaker, especially if there is background noise.

[trishume]

The thing is there is no way of knowing if humans are more efficient at chess than we are. Chess grand masters may be subconsciously doing a game tree search.
The reason chess does not use learning algorithms is that it is so straightforward that they are not needed.

The other thing is reliability. The best chess A.Is can beat the best grand masters nearly 100% of the time, whereas the best speech recognition is only 90% at best.

And speech rec is easy compared to things like object recognition in different angles and lighting conditions with a large database of objects. Again, the kind of thing 6 year olds can do in their sleep but the best algorithms are complete crap at.