Just a little disscussion. Do you think HTM's (Hierarchial Temporal Memory) will beable to take game AI to a new level? If you never read up on HTM's you can read here. Also if you read the book On Intelligence you will find extencive coverage of HTM's. I can see it taking AI and gaming all together to a new level. The only thing is getting it implemented properly. Now with more powerful hardware out there like the xbox360 a entire core can be devoted to AI. So you can only imagine the possibilies till it is implemented. What is your take on HTM's? For those of you who don't have time. The general gist of HTM's is a memory architecture that is based off a few algorithms to build a hierarchial memory system that simulates the brains neo cortex. Which controls all cognitive functions.

I see nothing in that white "paper" that have any credibility whatsoever. Its a big collection of unsupported claims. Actually, the only reference is the book the same guy wrote. Thats not serious.

While the authors credibility is questionable since he gives no other references than himself, the idea might have it's merits. I mean someone has to be first to put an idea to paper, and neural networks are a similar odd automagic system that gives surprisingly useful results if done right. Didn't have time to read the whole thing, but I probably will. But I would take it with a pound of salt, since the author time and again tries to tell the reader how HTM's are the best thing since DVD porn, and then points to his own book as the only refernce.

Yes, I believe a MCB (Magical Consciousness Box) would take Game AI to a new level. But I'm not holding my breath.

It is not totally bogus. This paper is the theory. The author did mention that in the begining of the white paper. He stated that there is no implementation in the paper only theory. He then stated that if you wish to get the mathematical paper on the algorithms you must purchase a license.

Quote: Original post by Wombah
... and neural networks are a similar odd automagic system that gives surprisingly useful results if done right.

SIGH. Again, neural networks are not an "odd automagic system". They are rather bad family of regression techniques. There are no evidences that they imitate the brain at all.

Quote: Original post by Wombah
But I would take it with a pound of salt, since the author time and again tries to tell the reader how HTM's are the best thing since DVD porn, and then points to his own book as the only refernce.

yes, exactly. And on the book website, he admits he have very few references in the book, because what he does it "so unique".

Quote: Original post by blewisjr
It is not totally bogus. This paper is the theory. The author did mention that in the begining of the white paper. He stated that there is no implementation in the paper only theory.

Who talked about implementation? Of course white papers are about theory. Implementation is just a way to get experimental data. But there is not much theory in the paper, only claims.

Quote: Original post by blewisjr
He then stated that if you wish to get the mathematical paper on the algorithms you must purchase a license.

/sarcastic Of course....

That doesnt add to his credibility.

Even if you write a paper about something totally new, its impossible that your idea is completely unrelated to anything in any field. You have to position your idea. Every claim that you make must be supported, then you show experimental data that confirms it. If you cant do that, then you have nothing of interest yet. That guy, a computer business man, not only claims that he invented the memory architecture of the future. He claims that he alone understand how the human brain works. Because he is just that clever.

It *may* not be totally bogus, but that guy surely takes every possible step to avoid being taken seriously. Including choosing a title that begins by "On ...".

Quote: Original post by blewisjr
It is not totally bogus. This paper is the theory. The author did mention that in the begining of the white paper. He stated that there is no implementation in the paper only theory. He then stated that if you wish to get the mathematical paper on the algorithms you must purchase a license.

Which puts him at the level of every scam artist and snake oil salesman on the Internet. Look, if he can show good, reproducible results, there are three dozen academic journals who would be very eager to make him famous, and three dozen Fortune 500 companies who would be very eager to make him rich. (He seems to be taking the latter route; more power to him.) It's great that he wrote a white paper on Yet Another Theory Of The Neocortex, and it seems to be an interesting one. But until he takes the normal steps to show people that he's not bullshitting them, he's not going to find much of an audience.

Quote: Original post by Steadtler

Quote: Original post by Wombah
... and neural networks are a similar odd automagic system that gives surprisingly useful results if done right.

SIGH. Again, neural networks are not an "odd automagic system". They are rather bad family of regression techniques. There are no evidences that they imitate the brain at all.

Sure, odd automagic system might be a faulty decription of neural networks. :)
What I meant was that it was when first proposed (I imagine) an odd way to go about things. If nothing else, that was how I saw it when someone first described them to me. It's not always trivial to see the connection between a number of neurons and actual results, and I'm thinking maybe HTM's might produce a similar form of 'emergent intelligence'.

Whether neural networks - or HTM's - actually model the brain accurately is of little or no consequence for the techniques viability (in fields other than actually modeling a brain that is ). Results are. So to reiterate I won't dismiss it entirely until I've seen an implementation, but I'll go heavy on the salt since the author seems flaky...

Quote: Original post by Wombah
Sure, odd automagic system might be a faulty decription of neural networks. :)
What I meant was that it was when first proposed (I imagine) an odd way to go about things. If nothing else, that was how I saw it when someone first described them to me. It's not always trivial to see the connection between a number of neurons and actual results, and I'm thinking maybe HTM's might produce a similar form of 'emergent intelligence'.

Artificial neural networks have always been and will always be a tool for parallel, piecewise regression. To what aim you bend that tool depends largely on how you represent the domain, how you combine the elements in your network and how you train your network. ANNs have never been an 'automagic' system or any other mystical pseudo-brain-like construct. They are what they are; a mathematical tool for manipulating data into a new space. It's merely the widespread lack of understanding of mathematics coupled with the spurious marketing claims made during the early '80s as to the potential of ANNs that has lead to a misconception that they are the panacea for AI research. It's much like the relationship between Fuzzy Logic and control theory. The latter is ground in nearly 100 years of solid mathematical and engineering research, while the former is a tool adept at solving a small class of control problems (or of being coupled with other methods to expand the class of problems). The former was overly hyped to the point of being claimed as the saviour of control, yet today exists only in research systems or in controllers for simple systems, such as thermostats. That's not to say it doesn't have merit on other problems, but there are often better tools for those jobs. Those working in control theory know the place of FL, just as those working in neuroscience or AI know the place of neural networks. They may have merit as a classifier, or for learning an arbitrary functional mapping, but there are other tools for the job that often perform better... (or are better instantiations of biological neural systems).

Cheers,

Timkin