An Algorithm Question

[CrazYmonkeY159]

Could you tell us more about your Neural Network, Training methods, How many neurons, Did you code it your self or are you running MATLAB?

Have you run any genetic algorithms through the network for training and testing?

So were using Mallet, open source java library for language processing. It works very well, however im looking for that little bit extra accuracy. I've tried the usual , Training methods can be specified, cross validation if you want, you can specify what you want your train/test split to be etc etc

so im basically just iterating(improving) on a current system by trial and error and im currently stuck on how to do this 'voting'

Whenever you classify something in mallet you get given its confidence (how good/bad the 'guess' is) and its classification, now if i have a few neural networks running and they are each trained using different (but proven) algorithms on a mixture of subsets of train/test data, in theory it should be more robust than a single classifier. That's what im aiming for but the paper i read was very vague on how the 'voting' was implemented. Hence me coming here

 

[Faultyboy]

Very interesting I hope it goes well for you.. I have read that it is better to keep the training quite specific to the network rather then using data that is irrelevant to the networks designed use. So if I wanted to train my network to identify cars, I would use random 'confidence' through out the network. Then I would insert all the parameters of a known car into the first layer of the network and specify what the result should have been. After a while of manually telling the network which answers are right it should be adjusting its confidence to provide a pretty good answer after a while... But if I go and run a bunch of other training data through the network not relating to identifying cars then it would change my confidence levels and not identify cars properly again.. I suppose this could be a problem for small networks with not as many neurons on the input layer. As the network gets larger and allows for more input parameters, you have the ability to train it for other purposes as you can input other data into the new input later neurons. But if 'other' data is put into the old already trained neurons it will spit out an incorrect result based on the wrong training.. In my eyes a neural network will never be able to recreate artificial intelligence as it's just a bunch of nested If Statements that have been trained to produce a certain result. Unless you can build a network that is so vast it allows for every possible type of input parameter there is in the universe and it will still need to be trained and what ever data it processes will still have to be inserted into the correct neuron. Very interesting but mind boggling topic..