So when you say you 'reversed' it; what do you mean? Using my example, (it is a good idea to provide such examples when asking for comments about such number related matters, btw ), do you mean that you went from this: 10,11 11,13 13,9 9,8 8,6 To something like this? 8,6 9,8 13,9 11,13 10,11 Or to something like this? 11, 10 13, 11 9, 13 8, 9 6, 8 And, how many training samples did you use?
Man, I'm a patient man, but you beat me soundly. Curious if anything constructive will come out of this, but if it does, it's entirely on you Wouldn't keep hopes up though. Garbage in. Garbage out. NN or not.
LOL...actually, it was my lack of patience that caused me to post. All of the prior responses were based upon not knowing critical facts about the problem. Very disturbing from my learned colleagues. I tutored other kids back in the day. My goal is always to show others how to approach the problem; how to think: and not just the solution. I'm hoping we can eventually get there. If the 'tutoree' is willing...we'll make it. If not, no hard feelings.
It was your second example: 8,6 9,8 13,9 11,13 10,11 And I can choose to use between 100 and 3000 training samples
Ok, without thinking much about it, I see a few issues. First, your belief that the order of the samples matters--for your setup. I've stated before that my thinking on machine learning is not based upon only academic papers and education. I think outside of the box. That said, I believe I've read where order will always matter. My opinion is that order should only matter when the NN is designed in a way such that the order matters. I don't offer papers or formulas etc. to support this, only a simple thought experiment. (The purported fact that order matters when it shouldn't, imo, is a flaw with NN's in general, imo; or a flaw in the particular platform, imo; or a flaw in design, imo.) Suppose you are a NN. I give you the following samples to learn: 3,9 4,16 5,25 6,36 7,49 Then, I ask you, what is the output when the input is 8? Then, I reverse the order, and have you relearn: 7,49 6,36 5,25 4,16 3,9 Then, I ask you, what is the output when the input is 8? Then, I ask you, did your answer change because the order was reversed. Then, I ask, why would someone even expect it to change? Why didn't it change?
This is exactly why I dislike so-called "IQ-tests". They simply assume too much and prescribe one specific way of understanding the patterns, while ignoring that the samples are just too few to extrapolate rules with high degree of certainty. Somehow, assuming one way of thinking, disregarding other ways of thinking and any kind of explanation, is considered "normal" and is something to measure against. It's easily verified that people training on such tests become better at them. So while intelligence is hard to improve in such a way, you can improve your understanding of what the examinators are getting at and score a bit higher due to that. Ditto with standard multiple choice tests and exams. Many ways to game such tests, so they must then be made to resist such attempts, containing options that could be right, but you need to select the "best answer" (according to latest dogma), etc.
That's exactly my line of thinking! I believe you should get the same answer when input is 8 - in each case.
So you agree that the order shouldn't matter. Why? because you don't tell the NN what the order is. You don't have the samples indexed (1,2,3,etc) and as an input. Nor do you have the date as an input. So, the NN should not be able to determine the order, let alone be aware that it has reversed. IOW, the stock data may "look" like it has been reversed, but to a simple NN--in this case, nothing has changed by reversing or otherwise changing the order of the samples...or at least nothing 'should' change.