Colin Kern - Revision history

Kernco: /* Journal */

2005-11-26T19:10:08Z

Journal

← Older revision		Revision as of 19:10, 26 November 2005
Line 15:		Line 15:

	== Journal ==		== Journal ==

			'''November 26, 2005'''

			I discovered an error in my script which explained why the error percentage was remaining around 50%. It was always doing only 1 iteration. I fixed it and tested some more. The error rate on the training data converges to 0%, but the network still performs around 50% error for the test data. The network seems to be learning the training data, but not adapting to recognize the pattern outside of the training data.

			'''November 24, 2005'''

			I decided to make a training file with equal amounts of positive and negative tests. The error percentage was around 50%, regardless of number of iterations.

			'''November 14-19, 2005'''

			I had a few tests this week, and I also needed to get some stuff ready for graduate school applications. I did some reading about training regiments, but I didn't get any useful information.

	'''November 13, 2005'''		'''November 13, 2005'''

Kernco: /* Journal */

2005-11-13T21:41:32Z

Journal

← Older revision		Revision as of 21:41, 13 November 2005
Line 15:		Line 15:

	== Journal ==		== Journal ==

			'''November 13, 2005'''

			I changed the training file so that it only included strings that were in the grammar. Now the network anwers close to 1 for every string. It seems like the network is only learning which answer is the most probable and not actually look at the strings.

	'''November 12, 2005'''		'''November 12, 2005'''

Kernco: /* Journal */

2005-11-12T20:03:41Z

Journal

← Older revision		Revision as of 20:03, 12 November 2005
Line 15:		Line 15:

	== Journal ==		== Journal ==

			'''November 12, 2005'''

			I still haven't gotten word of the extra functions being installed on Cairo. In response to some suggestions in class, I changed the training and test files so that it alternates putting each string in either the test group or the training group (I still want to test with strings not in the training group). The network still maxes out at approx. 19% errors. I looked at what the network was responding and discovered that almost all of the network's output was in the range of 0.00 - 0.09, and all of it was less thatn 0.3. This means that the network was rejecting every string, and 19% of them happened to be in the grammar.

			'''November 10, 2005'''

			I found a problem with the Octave installation on Cairo. There are some additional functions that aren't installed. I put a tarball for the admins to use so that I can start running things on Cairo. While I'm not parallelizing the functions, I can still run multiple tests simultaneously, which will speed things up.

	'''November 9, 2005'''		'''November 9, 2005'''

Kernco: /* Journal */

2005-11-09T18:26:33Z

Journal

← Older revision		Revision as of 18:26, 9 November 2005
Line 15:		Line 15:

	== Journal ==		== Journal ==
			'''November 9, 2005'''

			The error rate for \|a\| = \|b\| seems to stop going down at about 20%. After that, more iterations don't seem to improve the performance. I will modify my script to display which strings were incorrectly identified and try to see what is going on.

			I noticed that while training, Octave only shows 50% CPU usage. Running the training script with time shows this ouput:
			real 123m25.420s
			user 23m1.108s
			sys 90m8.374s

			I find it odd that so much of the time was in system mode.

			Parallelizing seems like it may be too much work to be worth it (unless my senior project will just be parallelizing the algorithm). Maybe I will use Cairo to just run multiple tests at the same time.

	'''October 25, 2005'''		'''October 25, 2005'''

Kernco: /* Journal */

2005-10-26T00:35:40Z

Journal

← Older revision		Revision as of 00:35, 26 October 2005
Line 18:		Line 18:

	I used the script to generate input data for the \|a\| = \|b\| language with maximum string length of 8. I started to experiment with training the net only using the string of length 4 or less, and then seeing how well it recognized the longer strings. I did 100 training iterations, recorded the error, then did 1000 iterations. The error with 1000 iterations was higher. I then ran the 100 iterations test multiple times and got different errors. I think that the weights are being initialized to random values, so I need to figure out how to save an untrained net and use it for each experiment.		I used the script to generate input data for the \|a\| = \|b\| language with maximum string length of 8. I started to experiment with training the net only using the string of length 4 or less, and then seeing how well it recognized the longer strings. I did 100 training iterations, recorded the error, then did 1000 iterations. The error with 1000 iterations was higher. I then ran the 100 iterations test multiple times and got different errors. I think that the weights are being initialized to random values, so I need to figure out how to save an untrained net and use it for each experiment.

			The error I got with 100 iterations was 46%, the error with 100 is 60%. It seems like after even 100 iterations the initial weights shouldn't make that big of a difference if the training is working.

	'''October 24, 2005'''		'''October 24, 2005'''

Kernco: /* Journal */

2005-10-26T00:31:51Z

Journal

← Older revision		Revision as of 00:31, 26 October 2005
Line 16:		Line 16:
	== Journal ==		== Journal ==
	'''October 25, 2005'''		'''October 25, 2005'''

			I used the script to generate input data for the \|a\| = \|b\| language with maximum string length of 8. I started to experiment with training the net only using the string of length 4 or less, and then seeing how well it recognized the longer strings. I did 100 training iterations, recorded the error, then did 1000 iterations. The error with 1000 iterations was higher. I then ran the 100 iterations test multiple times and got different errors. I think that the weights are being initialized to random values, so I need to figure out how to save an untrained net and use it for each experiment.

			'''October 24, 2005'''

	I wrote a Perl script that I give a maximum length and the number of symbols in an alphabet. The script will output all possible strings in the input style I am using for my neural nets.		I wrote a Perl script that I give a maximum length and the number of symbols in an alphabet. The script will output all possible strings in the input style I am using for my neural nets.

Kernco: /* Journal */

2005-10-26T00:03:47Z

Journal

← Older revision		Revision as of 00:03, 26 October 2005
Line 17:		Line 17:
	'''October 25, 2005'''		'''October 25, 2005'''

	I~~'m trying to write~~ a Perl script that I ~~can~~ give a maximum length and the number of symbols in an alphabet. The script will output all possible strings in the input style I am using for my neural nets.		I wrote a Perl script that I give a maximum length and the number of symbols in an alphabet. The script will output all possible strings in the input style I am using for my neural nets.

	~~I'm having some trouble getting it to work. It seems like the number of nested loops I need depends on the maximum length of the string.~~

	~~(Later that day)~~
	~~I have solved the problem~~.

	'''October 22, 2005'''		'''October 22, 2005'''

Kernco: /* Journal */

2005-10-26T00:02:25Z

Journal

← Older revision		Revision as of 00:02, 26 October 2005
Line 20:		Line 20:

	I'm having some trouble getting it to work. It seems like the number of nested loops I need depends on the maximum length of the string.		I'm having some trouble getting it to work. It seems like the number of nested loops I need depends on the maximum length of the string.

			(Later that day)
			I have solved the problem.

	'''October 22, 2005'''		'''October 22, 2005'''

Kernco: /* Journal */

2005-10-25T18:58:24Z

Journal

← Older revision		Revision as of 18:58, 25 October 2005
Line 23:		Line 23:
	'''October 22, 2005'''		'''October 22, 2005'''

	Using Netlab, I created a 2-layer perceptron with 4 input nodes, 4 hidden nodes, and 1 output node. I trained it with all possible strings in the grammar where \|a\| = \|b\|. After 100 iterations, the net correctly recognized 87% of the strings. After 1000 iterations, the percentage increased to 97%.		Using Netlab, I created a 2-layer perceptron with 4 input nodes, 4 hidden nodes, and 1 output node. I trained it with all possible strings (max length 4) in the grammar where \|a\| = \|b\|. After 100 iterations, the net correctly recognized 87% of the strings. After 1000 iterations, the percentage increased to 97%.

	Since the output node returned a floating point number, I rounded the values to determine the output on a "1 = accept, 0 = reject" scale. These decimals were generally between 0 and 1, however some were very small negative decimals, which rounded to 0. I didn't see any above 1.		Since the output node returned a floating point number, I rounded the values to determine the output on a "1 = accept, 0 = reject" scale. These decimals were generally between 0 and 1, however some were very small negative decimals, which rounded to 0. I didn't see any above 1.

Kernco: /* Journal */

2005-10-25T18:57:58Z

Journal

← Older revision		Revision as of 18:57, 25 October 2005
Line 15:		Line 15:

	== Journal ==		== Journal ==
			'''October 25, 2005'''

			I'm trying to write a Perl script that I can give a maximum length and the number of symbols in an alphabet. The script will output all possible strings in the input style I am using for my neural nets.

			I'm having some trouble getting it to work. It seems like the number of nested loops I need depends on the maximum length of the string.

			'''October 22, 2005'''

			Using Netlab, I created a 2-layer perceptron with 4 input nodes, 4 hidden nodes, and 1 output node. I trained it with all possible strings in the grammar where \|a\| = \|b\|. After 100 iterations, the net correctly recognized 87% of the strings. After 1000 iterations, the percentage increased to 97%.

			Since the output node returned a floating point number, I rounded the values to determine the output on a "1 = accept, 0 = reject" scale. These decimals were generally between 0 and 1, however some were very small negative decimals, which rounded to 0. I didn't see any above 1.

	'''October 17, 2005'''		'''October 17, 2005'''