I have recently begun studying neural networks and AI techniques and have a couple of general questions which I would appreciate some more experienced views on. I am not looking for exact answers to my questions just some plausible possibilities if possible. 1- If you added a second hidden layer to an MLP and this caused the network to perform fractionally worse what are the possible reasons for this. 2- If the training iterations were increased for the same network and this caused it to drastically decrease in performance could this be due to overtraining which means the network is not generalised enough to recognise patterns not in the training set. 3- If the network was designed to recognise images and there were 3 different resolutions for the images say low, medium and high I would have expected the high resolution images to perform better. However my tests have shown that this is not the case and no pattern could be seen. Scoring the network out of 100% I got the following results, low resolution 85%, medium resolution 83% and high resolution 84%. Could anybody give me a plausible reason for this? I thought the more information the more likely that the images could be distinguished from each other. 4- If you had two knowledge-based systems one with crisp rules and one with fuzzy rules and both performed perfectly then is the only advantage to the fuzzy rules system that it can give a degree of probability rather than a simple yes or no and is there any advantage in such a case for using the crisp rules. 5-Could any one give me a couple of practical advantages and disadvantages to adopting AI techniques in general. Any opinions greatly received Regards Elna

Quote: Original post by elna
1- If you added a second hidden layer to an MLP and this caused the network to perform fractionally worse what are the possible reasons for this.

Training becomes more complex with added hidden layers. The system may overfit due to the extra parameters, or simply fit poorly due to an inadequate training algorithm.

Quote: Original post by elna
2- If the training iterations were increased for the same network and this caused it to drastically decrease in performance could this be due to overtraining which means the network is not generalised enough to recognise patterns not in the training set.

This is not a question, but taken with question 1, makes me think that you may want to read the Usenet comp.ai.neural FAQ, which can be found at:

http://www.faqs.org/faqs/ai-faq/neural-nets/part1/

Quote: Original post by elna
3- If the network was designed to recognise images and there were 3 different resolutions for the images say low, medium and high I would have expected the high resolution images to perform better. However my tests have shown that this is not the case and no pattern could be seen. Scoring the network out of 100% I got the following results, low resolution 85%, medium resolution 83% and high resolution 84%. Could anybody give me a plausible reason for this? I thought the more information the more likely that the images could be distinguished from each other.

Your last sentence expresses a common misconception about decision making. More information is not always beneficial. Judicious selection of inputs is an important part of real-world modeling. Search for information on topics such as subset selection, feature selection, feature detectors, data reduction, etc.

Quote: Original post by elna
4- If you had two knowledge-based systems one with crisp rules and one with fuzzy rules and both performed perfectly then is the only advantage to the fuzzy rules system that it can give a degree of probability rather than a simple yes or no and is there any advantage in such a case for using the crisp rules.

First, fuzziness is not probability- it is the ability to represent degrees of truth. See the Usenet comp.ai.fuzzy FAQ to get started, which can be found at:

http://www-2.cs.cmu.edu/Groups/AI/html/faqs/ai/fuzzy/part1/faq.html

Second, if two systems "performed perfectly", then I don't see how one could have an advantage over the other, except in terms of size, cost, maintainability, etc.

Quote: Original post by elna
5-Could any one give me a couple of practical advantages and disadvantages to adopting AI techniques in general.

You will need to specify some other techniques with which to compare A.I. techniques- a thing can have an advantage only relative to some other thing.

-Predictor
http://will.dwinnell.com


[Edited by - Predictor on September 9, 2004 1:59:12 PM]