Discussion in 'Intelligence & Machines' started by Eidolan, Oct 30, 2008.
Ether way will be posible but short lived an becom extinct.!!!
Log in or Sign up to hide all adverts.
the human, the meat... robot ?
Connection machines (often called neural networks but I strongly dislike that misleading term) LEARN how to solve problems similar to those the were taught with. For example set of old bank loan applications (data about the person applying) and the subsequent late payment and default data on each. Then they were more accurate at predicting which new loan application to grant than the bank's human loan officers were.
They can learn to solve other problems and the weight of the various internal connections they have developed in learning phase can be examined by humans. Many times the human can not understand how the connection machine "thinks" better than he can to solve the problems.
To give a second example, many continuous fluid flow production process, such as paper or modern beer making have about a thousand variables, like the pH at 50 points along the flow, 50 temperature points in it, the flow rate, the drying rate, the air flow used to drying, how much bleach or hops to add and where, etc., etc. that determine the final quality of the paper or beer.
Typically some old timer, a master paper maker or brewer, had a "feel" for what needed to be adjusted as the inputs varied, but he could not tell you how he knew or what he knew and when he died, the company was in big trouble. Thus they instrumented everything they could and feed the data into a connection machine and told it the quality of the paper or beer that set of variables had produced. After a year or two of this training, the connections machine began to control these thousand or so variables but did not use exactly the same inputs the master brewer or paper maker did. Both, for example, tasted the flow at various points, but taste is hard to describe and instrument. None the less the output quality became high and more consistent. No human can control 1000 variables as well as the connection machines can and certainly can not understand how it is solving the problem.
SUMMARY: Some forms of AI are programmed by human to learn from examples, not programmed to solve some specific problems the human knows how to solve and could write code to do so. When the human does not know how to create code that will solve the problem, this is a good approach.
YOU ignorantly assume AI must be programed to solve problems, but that ceased to be the case at least 25 years ago.
The fundamental problem isn't that computers CANNOT learn, its that they cannot do it a fraction as well as we can.
The human brain has around 50-100 billion neurons, each neuron is many more times powerful then a single transistor, and then we have 1000 trillion synaptic connections.
I don't think there is any sort of electronic construct that actually rivals that. I don't think that the internet itself rivals it.
100 Billion Neurons with each having 7000 synaptic connections. Think of those connections as program breaks or steps where the logic can be branched out depending on the logic. So, now you have 100 Billion arm processors or smaller each doing 7000 steps. What is so difficult about it?
km, I went into this whole kind of amateur mini study of what would be needed for a computer to emulate the human mind.
The one big point is that it needs 1. emotions, 2. a subconsciouss that defines these emotions.
I don't mean emotions like love or hate, I mean purely, good vs. bad.
The current way for them to "think" is by having some sort of KNOWN algorithm based off of efficiency and it will make changes to the machine based on the idea of increasing efficiency.
The problem with this is that it is not truely learning. The reason is because the human mind is different because it can change in any imaginable way. Even the way we learn is subject to change. But since the way the computer neural network learns is governed by an algorithm which it knows. And because of this the fundamental way it learns cannot be changed.
What you need is a subconsciouss. A secondary machine which the primary Neural network cannot access, change, or alter. The secondary machine/subconsciouss will decide based on a very complex algorithm what is good vs. bad. It will then tell the primary machine if what it did was good or bad.
Similar to that game of hot vs. cold where someone hides an object and they tell you if you are getting hotter or colder. If you knew right from the start where the object was then there would be no point to it. But by having a "subconsciouss" that does know but cannot tell you then you can actually learn.
So effectively what you get is the primary neural network performing actions and then feeling an "emotion" based on that action whether good or bad. Since it cannot know exactly what makes it happier or sadder it will have to construct its own algorithm to explain the emotions. This will work until it finally does something that makes it sad despite it's created algorithm thinking it should make it happy, then it has to revise or destroy it. This is why you want as complex an algorithm as possible for the subconsciouss, that way it will take a long time for it to figure out the actual algorithm and "cheat" its way through the learning process.
This could also be an explanation for humanity's sentience (relative to homosapiens), a larger brain with far more neurons then necessary (at the time when it first evolved) means there is more room for a secondary conscioussness, a less dominant subconsciouss to form and become more complicated and advanced.
Please Register or Log in to view the hidden image!
Saturday Morning Breakfast Cereal
Please Register or Log in to view the hidden image!
Why do you say that, when in post 84 there are several examples of them learning to solve problems and control complex processes even better than humans can?
What you say is true ONLY of computers that humans supply with their processing algorithm. Perhaps you do not realize there is a whole class of computer which are supplied only with a learning algorithm (and some sample data to learn from). Did you read post 84? It seems not or your understanding is very limited.
"Several examples" can be considered as statistically relevent as a thousand person sample size for a statistic for a 300+ million person society.
Can they learn to do something BETTER, possibly. But the fact is that comparing neural networks to human beings is in a way apples to oranges. Machines work in mathmatics. Humans work in reality. And considering that the whole idea of mathmatics is to quantify, communicate, and describe reality in a similar way to how language is meant to communicate thoughts and ideas. Of course machines would be better to do something mathmatically then human beings.
The advantage of human beings are the flexibility allowed by the human mind. The ability to not only communicate ideas, but communicate tens of billions, even trillions of ideas to one another is incredible.
And in case you haven't read my post I actually meant the learning algorithm. And in case you were too narrow minded to consider that then read it over again and you will realize the meaning of it. That so long as the machine knows and understands exactly how it learns by having ready access to the algorithm then its ability to learn is hindered.
Billy T, you need to learn how to practice better science. What the hell is "...it seems not or your understanding is very limited."?
Your either saying that either I did not listen to you, or that I am wrong. You left no room in that statement for the possibility that you are wrong.
Out of the trillions of jobs, tasks, processes, etc... out there the human mind can accomplish far more of them then a computer can.
And quite frankly, a neural network computer constructed like a supercomputer across rows of interconnected computers arranged like blades inside of a large facility fails to impress me. A machine whose size, weight, energy consumption, and waste is orders of magnitude greater then the human brain fails to impress me.
Yah, sure you made a neural network that can play the stock market better then I can, but the fact is that: 1.) It is many, many, many times more inefficient, larger, heavier, and bulkier then I am. And 2.) While I may not be able to play the stock market better then it can I can easly drive a car, play soccer, and reason better then it can.
In 1828, Fredrich Wohler synthesized urea. The significance of this is that urea is unquestionably an organic compound (present in urine). Prior to its synthesis, no one had been able to create organic compounds synthetically leading virtually everyone to hold the view that "life" was somehow different from other chemical compounds. The theory was call "vitalism".
The argument that there will never be artificial intelligence sound like another form of vitalism.
Not true at all. A great many AI's have genetic learning algorithms that make possible the synthesis of entirely new reactions to stimuli.
I think you might be overestimating the majesty and uniqueness of emotion. It's just a drive. Machines can be given drives as well (i.e. avoid damage, avoid dark areas, seek a charger.)
To Fedr808 and others who know nothing about connection computers (called “neural networks” by most) here is a greatly over simplified example:
Assume a three layer machine. (It has been proven that it can do anything a machine with more layers can.) and that:
The input layer and the internal layer each have 26 nodes. Input modes 1 thru 26 and internal layer nodes A thru Z. (Why I chose 26, but it is good to have one input node for each fact known, even if it seems to humans not to be of much importance. (If that is true the connection machine will learn that and give near zero weight to the connections between the input node corresponding to that fact and all the internal nodes.)
Let’s assume there are only two output nodes correspond to a binary decision to be made. For example grant or not to grant a loan application.
Some of the inputs will be continuous variables (for example applicants age, perhaps coded a voltage between -1 and 1 Volts) and others binary (Male vs Female, Married or not, etc. perhaps coded as either -0.5V or + 5V) There can be strong negative feedback between the two output nodes. This forces one to be nearly +1 and the other to be -1. Or the human user can just take the more positive one as the result.
Initially all connection weights are random strengths between -1 and 1. For example initially the weight on input node 12 to inter mediate node M may be -0.7452 and from 12 to K may be +0.3867 etc.
Then a simple learning rule could be: If output on a learning trial is correct, then all of the positive connections are slightly increased and negative ones made slightly more negative. The next learning trial may further increase the connection strength of some connection and reduce some of the weights that were increased by the prior trial. After many learning trial have been made (all with different data samples) the connection weight rarely change much, if at all with the next dozen or so trials. Then the connection machine computer can be used to make judgment calls on cases it has never before seen.
I.e. it can take the available data about a new loan applicant and decide if he should get the loan he is requesting. If connection machines, trained up on prior loan payment results, had be used one could be nearly sure that the current financial crisis would not exist as they would never grant loans to people very unlikely to pay back the loan. They don’t suffer from greed for loan placement fee or the smile of a pretty girl, etc. They just use loan repayment histories to decide. Note no human programed the machine to make these decisions - It learned how to decide. The very same machine (with different inputs and historic data) could learn to give advice to person 1 about marrying person 2. etc. THE MACHINE LEARNS, how to decide.
This is not just obvious, but about 25 years ago was tested by some banks with the results that the connection machine's bad loan rate was lower than human evaluated loan applications but that was too hard on the egos of the human loan decision makers so was discontinued as they did other things for the bank, not just decide about loans.
PS to fedr808:This serves as a reply to your post 91, but I need to add that connection machine are typically very small compared to a lap top. In fact 20 years ago you could buy one on a chip.
Your assumption about size, weight and power needs is very wrong.
True that they only can be used for the problem you trained them up to solve. I.e. they lack human flexibility, but are extremely fast. (Millions of times faster than any human at reaching a decision and that decision can change as more data is given just as fast.) There is no long set of steps to follow as common in programmed computers. The result is essentially immediately available - sort of like a resistor summing network. In fact after trained, they can be replaced by a simple resister summing network. For this, their very light weight and immunity to radiation jamming etc. they are used by the military.
Another extremely important point is they can learn to solve problems that no human knows how to create a solution algorithm for. I gave some examples of this in post 84 with complex industrial processes like paper making and beer brewing where some batches turn out good and others not so good and no one knows why.
Your idea that connection machines "work in mathematics" is also wrong. They work in volts and current flows much like the resister network that can replace any trained machine. For convenience they are almost always designed in a regular digital computer's simulation of the physical connection machine as then there is no soldering etc. to do. The changing connection weights while they are learning can be automated; no need to replace each resistor making the node-to-node connections with another on every learning trial.
I am sure they NEVER have been applied to a mathematical problem as you seem to be assuming is their use. What would be the "training set?"
Wow Billy T, your ego is so massive and so dense that there is an iminent threat of it's collapse.
But more to the point. Well actually, you haven't really made a real point. My original post was about what needs to happen to create a computer with actual intelligence, with higher level heuristic ability.
There is a reason they are called "Artificial Neural Network" or ANN, neural network is merely what you have, not what a computer has.
ANN's do not possess heuristic abilities comparable to a human's. That is the essential problem.
An ANN is really nothing more then a network with so many nodes that it is impractical to manually adjust the weights in order to gain the desired output of the ANN, the learning algorithm commonly used is the backpropagation algorithm.
The problem remains that the network itself arguably is not learning so much as it is merely adjusting and tinkering. Learning implies understanding of the actual problem and the self awareness to know why it is solving the problem. An ANN lacks both of these.
It is merely adjusting the weights of the "neurons" until it can obtain the desired result.
Tell me Billy T, what would happen to an ANN if I were to give it a problem and say "figure it out". An ANN's ability to figure out a problem, understand it, create a method to solve it, solve it, and then critically analyze IT'S own solution to figure out if it was the best one.
I have no idea what kind of stunt you are trying to pull Billy T, but thus far you have done little other then to state common knowledge.
No true. Here is your original post:
My "real point" as you say, in several posts, was just to show your post 85 claim is false or at least overstated. In many limited fields they learn to solve problems better than humans do,even some problem humans can not learn how to solve well enough to write an algorithm for conventional digitial computers.
You now understand that*, so you have been moving the goal posts - Now demanding flexibility, heuristic ability, creativity, etc.
I have several times agreed that connection machines lack these human abilities, but had to occasionally correct new errors you made, such refuting you claims about how large, heavy, and power hungry these machines were compared to a human brain** by point out that 20 years ago you could buy a connection machine on a chip! Also you falsely claimed they were only good for "mathematical problems" and I refuted that too with examples. I even went out on a limb to state that connection machines NEVER do mathematical problems as there is no training set data for them to learn with.
I am not seeking to argue with you - I am just doing what I often do - Correct nonsense when I read it in a post. For years, I have held the unofficial title as "the Sheriff of Nonsense" and I have "arrested" dozens of others posting nonsense by clear refutation of it.
* You even admitted they could learn to be better than humans is some applications a few post back.
** The human brain may use 1/3 of the body's total energy when you are sitting at a desk thinking. A modern low-energy chip uses hundreds of times less energy. It also weight about 1000 times less and is at least 500 times smaller in volume. - Three strikes and you are OUT, is the normal rule.
Of course it would still exist. It just takes in data sets (provided by humans) and produces data (looked at by humans.) If the goal is "make me lots of money in the next year so I can get a promotion" then that data might well be used to give loans to poor credit risks - because the money made in the next year would (for the person deciding how to use the data) outweigh the risks in the next ten years of defaults.
Neural networks are great tools to analyze fuzzy data. But they are just that - tools. They do not decide economic strategies or investment options.
I agree that the current economic crisis was largely caused by greedy humans; but I said that it would not have been likely to occur IF connection machines, well trained on prior loan applications and the corresponding repayment records, had made the "lend" or "not lend" decisions.
We agree connections machines did not make those critical decisions, humans did, so now we must live with troubles excessive human greed can cause.
To distort an expression about war: "Economics it too important to be left to humans." Please Register or Log in to view the hidden image!
Well by your definition of "learning" any system that can alter itself in such a way to increase the efficiency/success of said system.
Lets emphasize the words "limited fields". So what if it can do a dozen jobs many times better then I can. I can potentially (and I stress that word) perform thousands, even millions of different tasks much better then it can, including constructing a better version of said ANN.
The fact is that the amount of tasks an individual person can potentially perform relatively well versus how many an individual ANN can potentially perform is many orders of magnitude greater.
Tell me, are ANN's self aware?
All that ANN's are at this point are machines that perform a task a few different ways, determines which configuration produces the best result, selecting that configuration and repeating.
If you delude yourself into thinking that this ability can really be considered artificial intelligence let alone evidence of significant learning then you need to learn what an ANN is.
The human brain contains far, far, far, far more nodes and far more connections then an ANN does.
I have not been moving the goal posts, merely rewording my point so that someone like you can grasp it.
And the fact is that you are still wrong about your point on the size of a chip.
If every node on an ANN is comparable to a real neuron then tell me, what exactly is the size of an ANN containing 50-100 billion nodes and 1000 trillion connections (which at this point may not even be possible to create that many connnections, but I don't know that off the top of my head).
The fact is that there exists no single chip the mass and dimensions of the brain with similar or less power demands that can best it in every single possible task that a brain can perform.
*Please, I use the term "learning" so I don't have to confuse you by using more accurate terms such as "adjusting" or "tinkering".
** Really? And can that chip perform heuristic thinking? Can it learn with comparable accuracy? Moreover, can that chip, on it's own with no help from any exterior sources best a human brain in every single possible task? Hell, that chip could not even design an improved version of itself better then a human brain could. A modern low energy chip is also complete and total crap compared to the total processing power of the human brain.
Or even go a step forward and make machines that analyze various inputs such as the growth of a company, change in stock price, equity, loans, and the like and determine if there is anything out of place with the company.
It wouldn't replace human investigators but it could help find the few corrupt companies out of a very large amount of good companies.
Separate names with a comma.