[theCPE]

I was very intrigued with what Hawkins described as the cat-dog problem specifically because of how much sense it made. However, I don’t think that the problem can be solved by merely using pattern recognition or HTMs.

I read an article somewhere that Numenta had tested their systems with a simplified cat-dog problem and it worked and therefore was looking very promising. However, unless I misread, the cats and dogs that were being used were stick figures.

If stick figures were used than specific strokes must had been assigned as being cat like or dog like by an observer. Also, the amount of data that must be understood and patterns found within is diminished immensely with stick figures. And the biggest problem, if specific strokes in stick figures are designated either cat like or dog like you are removing from the problem the biggest issue in nature, variance. There is an infinite amount of variance between cats and likewise with dogs. Also, there are thousands of angles and perspectives to view the cat or dog from adding even more variance to the dilemma. This leads to my main point.

How with so much variance available can pattern recognition be the source of success? If you are dealing with a very small domain than finding patterns is easy, but finding patterns from a nearly infinite domain seems daunting, especially if the success of the system is expected to match the success of humans. If the success is only desired to be greater than 50% than that is entirely different. However, when was the last time you met a person that could only determine if it were a cat or dog more than 50% of the time? Or when was the last time a person looked at a futuristic silly looking chair (ones from MIB come to mind) that had no resemblance to other chairs and needed multiple guesses to get it right?

While pattern recognition and the HTMs might be more successful than current systems at object recognition, I have doubts if it will result in human level object recognition. We see things all the time that in no way resemble anything we've seen before, but we know what it is....through function recognition. Basically, a system only designated to find patterns in how things look won't work nearly as well at object recognition as humans. A human brain isn't just macthing patterns to what it is seeing now from previous objects, it is using knowledge of its world to determine what functionality it might possess, what is the objects environment, and how could the object be manipulated to be useless. These ofcourse are just guesses. Anyway, a few thoughts.

[theCPE]

I think a good analogy of what I am trying to suggest is the following:

Given a word you have never seen before and asked to define it is difficult. Besides breaking down roots of the word and making guesses there is little hope to get it right. However, if you are given a word within a sentence it is far easy to define the word thanks to context clues and you will be right FAR more often.

The human brain doesn't just see "words", it sees "sentences", I am afraid that the HTMs are only seeing the "word".

Any thoughts?

[intelligent robot]

The issues you pointed out are actually the same criticisms that Hawkins makes of traditional AI systems; and they are issues that HTMs are intended to solve.

In fact, I believe something quite similar to your second post appears in the book on intelligence. I encourage you to read this book if you haven't already.

About the reduction of information in stick figures, you are right; the models that HTMs have been tested on so far are very simplified - not because HTMs are inherently unable to deal with complex information, but because the computational power of current test machines is very low.

In the book, Hawkins expresses hope that we might be able to circumvent this problem by an easier way than trying to build ever faster computers. He thinks that it might be possible to develop cortical algorithms directly into hardware. I think this is highly possible and will allow human-level thinking much earlier than that foreseen by current computing power trends.

[theCPE]

I read Hawkin's book about a year ago, but I don't recall the "context clue" analogy. That is where I was introduced to the cat-dog problem and I am pretty sure that Hawkins in the book disagrees with computing power being the dilemma with attempts at AI.

Basically I think his logic went: A human in a fraction of a second can look at a picture and say cat or dog. Only 100 neurons in series(aprox) could be triggered in that time frame, yet there is currently no way to have a computer determine cat or dog within 100 steps after being given a picture.

Basically the logic suggests computing power is not the problem.

If in his book he eludes to the context clues like my previous post I must have forgotten that portion. But I believe that makes the most sense.

Basically creating a partial brain (the HTMs for object recognition) will not create the results a whole human brain can with object recognition. The best that could be hoped for is better than 50% accuracy.

[intelligent robot]

[theCPE]

.....1 neuron represents 100 sequential instructions from a computer.....

are we discussing the same book?

[theCPE]

If you look over pages 66-68 of his book you will see what I am talking about.

Computing power is NOT the problem with computers and being intelligent or performing object recognition, the thought experiment is explained in much more detail on those pages than my brief description.


Further, my "context clues" analogy can be summed up with this example.

When I look at something I don't just see IT, I see the entire environment it is surrounded by. I see its location, I see its orientation, I see near by objects. All of which are immediately triggering neurons and simultaneously causing pattern recognition to access "memory". Further, also stored in my brain is knowledge or facts about my world. Knowledge of gravity, knowledge of physics, general knowledge of how the world works. The brain probably works in a massive parallel heirarchial system of memory (all conjecture). All of these stimulae are necessary for me to determine what I am looking at, how to use it, what is its purpose, what sounds I produce with my vocal cords that represent its name etc.

While using strictly pattern recognition on visual stimulae may provide decent results eventually in guessing dog or cat, without a full brain with full world knowledge it wont have results like that of a human brain.

Ie the context clues.

Anyway....was hoping someone within numenta frequented these forums.

[Tom11]

I haven't read the book yet, but anyone who knows anything about the brain knows that any sort of software recreating _genuine_ intellience would indeed require very large amounts of memory and processing power.
You're either ignorant of the brain or are being disingenuous.

Also, no one is _born_ with any facts or knowledge of physics, or anything else. It all must be learned.

A large HTM will learn about physics, gravity, and other things (or causes) by observing the behaviors of all these causes (which will often be sampled by other senses beyond just sight), and how they relate and interact with one another over time. This is how an HTM, over time, builds a model of its world, in the same way a biological brain does.

Also, why do you insist on talking to someone from Numenta when there are many others who frequent this forum with much more knowledge about this than I have, 'Intelligent Robot' being one of them. They're very busy working on their HTM platform right now, so that they can get it out to the community before the end of this quarter.

I suggest you wait a couple of months or so, and read their whitepaper at www.numenta.com, in the meantime.

[theCPE]

[Tom11]

[quote="theCPE"]

[theCPE]

Holy crap. Ok, big mistake in getting my hopes up about discussing theories and science behind object recognition.

Between people here that completely misread or misinterpreted Hawkin's own thought experiments and examples in his book and others that have to immediately take things personally and call people ignorant for making conjectures about oversights that might exist in current theory I guess having thought provoking discussion here is too much to ask!!

Obviously everyone here knows the answers to everything and has their own HTM running flawlessly in their basement pointing out any and all things with 100% accuracy.

But just to beat a dead horse a bit more, Hawkins himself repeatedly in his book points out that computational power IS NOT THE LIMITATION TO AI!

And because I didn't come here to engage in a pissing contest with you Tom, I will concede that you are obviously right I am entirely ignorant of everything and your dad could beat up mine.

[intelligent robot]

No need to draw conclusions so fast, TheCPE.

From what I understand, you are trying to argue that:
1) Object recognition is much more complicated than Hawkins thinks, and
2) Hawkins is mistaken in the requirements needed for accurate brain simulation.

Both of these arguments are wrong, in my opinion.
First of all, Hawkins acknowledges the fact that there is much, much more detail to the workings of the brain than we could ever hope to gather into one consistent algorithm. He is aware that HTMs are at best a simplistic representation.
However, you should note that simplistic representations are not always inaccurate. Very simple algorithms have been used before to represent complex phenomena. I encourage you to read the book A New Kind of Science by Stephen Wolfram; a broad discussion of this fact is given.

In fact, the leading consensus in intelligence research is that despite the almost infinite level of complexity in the brain, a simple algorithm may be able to recreate its complex processes. Proponents of HTM technology (myself included) believe in this point of view.

Also, Hawkins mentioned the '100-step rule', to draw a comparison between the capabilities of computers and neurons.
Just so I could make an intelligent argument, I re-read the book and, after this argument, he points out that the mechanism of the brain is 'to retrieve patterns from memory'. He does not say that we have a small amount of memory and/or processing power. In fact, quite the contrary; we all realize that the memory-prediction framework will require huge amounts of memory to even approach the capabilities of human brains. However, with current trends, we remain optimistic that such a goal can be realized in a very short amount of time; perhaps a decade or two in the future.

In the meantime, HTMs with lesser memory requirements (that will be out before this year is ended) will be used to perform many intelligent tasks, such as driving our cars or coordinating our resources.

If I understood you wrong, please say so.

[theCPE]

[Tom11]

[intelligent robot]

TheCPE, I believe that you might have misunderstood Hawkins debate about processing power and intelligence. He claims that traditional AI researchers will fail, because they ignore the algorithm, instead claiming that the only boundary to their work is processing power.
He claims that this is a fallacy because they largely ignore the hidden workings of the brain. I think we both agree on Hawkins' viewpoint here.

However, nowhere in the book does Hawkins claim that it is possible, for example, to simulate a brain on a 486. That is largely absurd.

You try to draw a boundary between memory and processing power, but in computer science they are both relatively isomorphic. To do lots of processing, you usually need lots of memory, and to use lots of memory, you ultimately need a lot of processing power to work on that memory. Memory is useless without being able to do operations on it (such as storing and retrieving), no?

So, to sum up what I grasped from the book, Hawkins argues that computational power is not the only boundary, and also not the most important boundary. The most important boundary is our use of that computational capacity. He holds that with the proper hardware it might even be possible to reach human-level amounts of intelligence in just a few years. With this I largely agree.