Sent a email to professor about our situation?

[dhye]

Hi guy,

Do you think we can finish the project? I am stuck and maybe we sent email to professor for help, such as that our code is worked for xor, but it is not work for the project. I don't see what is wrong with the code. 

@Ricado

Write up an email example for help and send to us. Can you do it in half hour? Give me a email or call when you are done.

--

Best,

Daihua Ye

College of Letter and Science,

University of California, Davis

B.S. Computer Science

dh...@ucdavis.edu | (408)-386-7053

My Portfolio

[Thomas Brounstein]

Hey,

I can be on campus where ever you guys need me all day tomorrow to finish this up if we can't get an extension.  When and where do you want me?

--
You received this message because you are subscribed to the Google Groups "AI_Connect" group.
To post to this group, send email to ai_co...@googlegroups.com.
To unsubscribe from this group, send email to ai_connect+...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/ai_connect?hl=en.

[Thomas Brounstein]

Also, going over the code to see what I can find.

[Daihua]

We possible go to the csif at noon. I often work at night, so i wake up late. I possible can come at 2pm in csif.

Sent from my iPhone

[Thomas Brounstein]

Got it.  I tend to go to bed earlier (and wake up earlier).  I'm looking at the code right now, though, and trying to piece together what's happening.  A first thought: are we over fitting the data?  We might be running the training data through too many times which could be causing problems.

Also, if we do implement standard deviation (I was sure it was in Java already!) the formula can be found here:

http://standard-deviation.appspot.com/ 

[Thomas Brounstein]

Also, it seems we have two layers of hidden nodes now.  What is the reason for this?

[Thomas Brounstein]

Another thought: it looks like the training weight is 0.5.  That's rather high, so perhaps it's causing the error?  What if we reduce it to 0.1 or 0.05?  Does that improve performance?

[dhye]

We don't have two layer hidden nodes.

[Thomas Brounstein]

I think I was looking at the wrong file at the time.

[dhye]

The training weight is less than 0.5. According to the formula: (float) Math.random()- 0.5f;

Math.random() is in 0 to 1. So (float) Math.random()- 0.5f is between -0.5 to 0.5.

[Thomas Brounstein]

I think we want our training weight to be a small, positive constant.  So just hold it at .1 or so.

[Thomas Brounstein]

Looking at the code, I'm somewhat confused on some key points.  Specifically, we need to run the method twice, once to train and once to test.  However, it seems that for this five fold validation that isn't the correct construction: we want to run the code once and have it select five random groups, and for each random group run "train" on the four other groups then "test" on the fifth group (i.e. pull one group aside, build the network with everything else, then test on the group we pulled aside, then repeats for each other group).  I'm not sure this fixes the problem on hand, but it is something to be aware of.

I think having a constant, small, positive scaling factor should also help.  I've been trying to parse through the code as well but I'm having trouble seeing anything critically wrong.  Unfortunately, I'm also too tired to continue tonight.  I'll meet up with you guys around noon tomorrow and hopefully we can get this working.

Tom

[dhye]

For part 2, we don't run the method twice, because the 5 fold are randomly generated and they will have different folds each time. For part 2, I am confuse about how to test with 5 fold. My understanding is to choose one of 5 fold and test with the other one. So I think I am wrong. 

For the small constant, small, positive number (0.1). it is still not working. The output is always 0.0.

[Ricardo Nunez]

Thats true, decreasing the constant always makes the output 0.0 in my tests. I've been trying to debug this since Friday night, still haven't found anything explicitly wrong in our implementation, but can meet today after 2:00 at csif if you guys want.

[Thomas Brounstein]

I've been looking through it most of last night and this morning and still don't see anything wrong.  Hell, I've been posting everything I see that *might* be wrong, and clearly none of that has actually been in error.  I'm having trouble actually running the code (due to me not knowing how to work with Java as well as I should--basically I'm having a hard time calling external files), so I can't check, but I have a quick question.  The output is always 0; do we know what the hidden nodes return, or what the weights on the hidden node to output node edges are?  I'm wondering if the error is between the input nodes and hidden nodes, or the hidden nodes and output nodes (or both).

Oh, one other thing.  Do we have a constant term in the implementation?  I don't remember seeing it, but I could be missing it.  That might be causing some of the problem (though I would be surprised if it fixed much).

I'll be at csif at 2:00.  Which room (aren't there multiple rooms in the csif?)

[dhye]

After, I run the code the Weight between the input nodes and hidden nodes is 0.1. They all are not updated. Also, the weight from hidden to output nodes are -859.7421. The program is use 0.1 as default weight.