@p6v53as435fe Kinda. lol, Actually more nodes would make it slower but more accurate. This is just a demo I made in Excel. It works but Excel is not designed for this. If I needed it for a real problem, I'd have a programmer do it in VB or something more suited to computation. I'm not a real programmer though I dabble. I just wanted to see if I could make it work. I did and I'm happy :)
@Marksman560 I'm not sure if you're just trolling, but on the off chance that you're serious....
The 2D aspect of this video is just the video. It's a 2D rendering of what's going on mathematically. The video rending does nothing to help the process. In fact it slows it down quite a bit. If I were going to describe the neural network I'd describe it as "single parametered" or "30 weight." Hope that helps.
You do not describe the problem, nor the solution, nor the process... this has zero educational value, and no technological tools are displayed.(excel??)
Showed like this., it is an unexplained graph, void of meaning or context.
I don't see how the title which have brought me here to watch is being given any credence in relation to the content.
@superdiza This wasn't posted for your benefit. You are no one important but I suspect that's not the first time you've heard that. This was posted so people who matter could see it and discuss it with me. You are nothing and probably too stupid to understand such things. What have you posted? Nothing. Troll.
@cybercab Hi, I dont mean to start trouble again but he does have a point :( Im completely new to Neural Networks and Genetic Algorithms like so many other people in the world, and sadly as I dont know much, it could help if you could describe what the problem is and what the "AI" is trying to achieve, is it attempting to be similar to the blue graph?
@EZbakeROFLcakes No problem. I don't mind answering legit questions. The AI is trying to find the relationship between many parameters (e.g people's height and weight). The AI part is revising the equation constantly to find a better fit to the graph. It's rather involved but Wikipedia has good info on the "gradient descent" approach.
A few questions: Of how many samples is the dataset? How many input features are there per sample? Are the 30 parameters(weights) you mentioned just in the input layer, or spread across the entire network? Do the nodes ultilize a binary activation threshold? Is the parametric or nonparametric? Stochastic? How do you avoid over/underfitting? Are you using minimum least squared error? Have you run the alg. on novel data, and how did it react? Ok..ok... done... thanx in advance. :)
I am using a stochastic approach though it sounds like you know more about this stuff than me. :) I'm using hyperbolic tangents. 30 weights but 10 data points. I avoid overfitting (bigger problem) but another trick. If it goes 100 iterations without improving it decreases the adjustment. It's still pretty sloppy but it gets to 99.999% after 1M trials. I tested it in Excel and it works but to actually use this you'd want to do it it VB or C+. Excel has too much overhead. This was just a test.
Nice work Cybercab. You mentioned a simple four step function with gradient decent, and this is obviously a nonlinear mapping of input features along an objective function, which leads me to believe you are likely using a single hidden layer backpropagating multilayer perceptron with a sigmoidal transfer function(possibly softmax?); however still I can not help but wonder why on Earth it would take so many millions of iterations for the algorithm to converge to a minimum of the error function.
Hi, Cybercab, it is very interesting work. Did you code in VBA to do a NN macro?
Could you please tell me, behing these blue graph, there is only one input (x) and one output (y)? I do lots of sales forecasting job. Could you please anser my questions and if possibly provide with hints of how to do something similar in Excel?
Yes, this is ALL in Excel. I used a macro as well. It's 1 math function and 4 steps repeated over and over. No real programming (other than macros). It's really slow in Excel but I'm not a programmer. Overnight it will run 1 million trials and get very accurate. I can send you a copy of the XLS file if you want.
cybercab, what exactly is being demonstrated here? What does the blue line represent and hence how is the red line learning to describe or copy or predict it?
Thanks. The blue line is the real relationship between X and Y. The red line is my guess. Then the NN alters the equation slightly and checks to see if that's better. If better then it keeps going. Over time, it literally learns and changes its approach. This is great for economics, engineering, or anything. Imagine understanding how 15 parameters affect your car's performance. A NN can tell you!
It's very impressive. I've not written a succesfull NN yet. I've had a lot better luck with GA's. I guess i was able to mentally envision what was happening with them better.
I guess it needs more nodes to learn it faster?
p6v53as435fe 9 months ago
@p6v53as435fe Kinda. lol, Actually more nodes would make it slower but more accurate. This is just a demo I made in Excel. It works but Excel is not designed for this. If I needed it for a real problem, I'd have a programmer do it in VB or something more suited to computation. I'm not a real programmer though I dabble. I just wanted to see if I could make it work. I did and I'm happy :)
cybercab 9 months ago
You take an input (x) to an output (y) so it isn't a multi-dimensional problem :P
Marksman560 1 year ago
@Marksman560 I'm not sure if you're just trolling, but on the off chance that you're serious....
The 2D aspect of this video is just the video. It's a 2D rendering of what's going on mathematically. The video rending does nothing to help the process. In fact it slows it down quite a bit. If I were going to describe the neural network I'd describe it as "single parametered" or "30 weight." Hope that helps.
cybercab 1 year ago
@cybercab
Aha. I was misreading your video-comment ('2D 30 weight neural network'). btw 'single-parametric' is a nice description for what I was saying :)
Marksman560 1 year ago
this is silly.
What's the algorithm? Excel sounds cool, but where' the cool stuff?
superdiza 2 years ago
This is silly? I'm not sure I understand what you mean.
You can look up the details on Wikipedia. Look for "Gradient descent"
cybercab 2 years ago
well, I did.
I still think this video is silly /pointless.
You do not describe the problem, nor the solution, nor the process... this has zero educational value, and no technological tools are displayed.(excel??)
Showed like this., it is an unexplained graph, void of meaning or context.
I don't see how the title which have brought me here to watch is being given any credence in relation to the content.
superdiza 2 years ago
@superdiza This wasn't posted for your benefit. You are no one important but I suspect that's not the first time you've heard that. This was posted so people who matter could see it and discuss it with me. You are nothing and probably too stupid to understand such things. What have you posted? Nothing. Troll.
cybercab 2 years ago
@cybercab so much ego, so many assumptions.
as a youtube user I have the right to comment and rate, and I did.
you have the "right" to call commentators names and feel better about your self.
have fun :)
superdiza 2 years ago
@superdiza "Rights?" What are you 14? You're a troll. Looking to make trouble. You've still never posted a video. I guess that sums you up. Nothing.
cybercab 1 year ago
@cybercab Hi, I dont mean to start trouble again but he does have a point :( Im completely new to Neural Networks and Genetic Algorithms like so many other people in the world, and sadly as I dont know much, it could help if you could describe what the problem is and what the "AI" is trying to achieve, is it attempting to be similar to the blue graph?
EZbakeROFLcakes 1 year ago
@EZbakeROFLcakes No problem. I don't mind answering legit questions. The AI is trying to find the relationship between many parameters (e.g people's height and weight). The AI part is revising the equation constantly to find a better fit to the graph. It's rather involved but Wikipedia has good info on the "gradient descent" approach.
cybercab 1 year ago
@cybercab Thank you :D
EZbakeROFLcakes 1 year ago
A few questions: Of how many samples is the dataset? How many input features are there per sample? Are the 30 parameters(weights) you mentioned just in the input layer, or spread across the entire network? Do the nodes ultilize a binary activation threshold? Is the parametric or nonparametric? Stochastic? How do you avoid over/underfitting? Are you using minimum least squared error? Have you run the alg. on novel data, and how did it react? Ok..ok... done... thanx in advance. :)
mrtransist0r 3 years ago
I am using a stochastic approach though it sounds like you know more about this stuff than me. :) I'm using hyperbolic tangents. 30 weights but 10 data points. I avoid overfitting (bigger problem) but another trick. If it goes 100 iterations without improving it decreases the adjustment. It's still pretty sloppy but it gets to 99.999% after 1M trials. I tested it in Excel and it works but to actually use this you'd want to do it it VB or C+. Excel has too much overhead. This was just a test.
cybercab 3 years ago
Nice work Cybercab. You mentioned a simple four step function with gradient decent, and this is obviously a nonlinear mapping of input features along an objective function, which leads me to believe you are likely using a single hidden layer backpropagating multilayer perceptron with a sigmoidal transfer function(possibly softmax?); however still I can not help but wonder why on Earth it would take so many millions of iterations for the algorithm to converge to a minimum of the error function.
mrtransist0r 3 years ago
Hi, Cybercab, it is very interesting work. Did you code in VBA to do a NN macro?
Could you please tell me, behing these blue graph, there is only one input (x) and one output (y)? I do lots of sales forecasting job. Could you please anser my questions and if possibly provide with hints of how to do something similar in Excel?
Thanks.
alexeymosco 3 years ago
Yes, this is ALL in Excel. I used a macro as well. It's 1 math function and 4 steps repeated over and over. No real programming (other than macros). It's really slow in Excel but I'm not a programmer. Overnight it will run 1 million trials and get very accurate. I can send you a copy of the XLS file if you want.
cybercab 3 years ago
Thank you, I would very much appreciate it.
alexeymosco 3 years ago
Hi Cybercab,
can I also have a copy of the spreadsheet file?
nikiforosc 3 years ago
may I have it too..please ?
liadon666 3 years ago
cybercab, what exactly is being demonstrated here? What does the blue line represent and hence how is the red line learning to describe or copy or predict it?
Very interesting btw.
talkingbookworm 4 years ago
Thanks. The blue line is the real relationship between X and Y. The red line is my guess. Then the NN alters the equation slightly and checks to see if that's better. If better then it keeps going. Over time, it literally learns and changes its approach. This is great for economics, engineering, or anything. Imagine understanding how 15 parameters affect your car's performance. A NN can tell you!
cybercab 4 years ago
i'm surprised that 30w acheived so low a resolution. what training method did you use?
damianpoirier 4 years ago
30w really only gives you 10 guesses. I could increase it but it gets really slow.
cybercab 4 years ago
It's very impressive. I've not written a succesfull NN yet. I've had a lot better luck with GA's. I guess i was able to mentally envision what was happening with them better.
damianpoirier 4 years ago
I'm using gradient descent methodology.
cybercab 4 years ago
Pretty neat.
And thanks for not posting the 6-hr version, I'm happy to take your word for it.
ddlsr 4 years ago
You're amazing. I had no idea Excel could do this!
Zarbod 4 years ago
Thanks Z. I'm subscribing to you!
cybercab 4 years ago