Page 1 of 3

training network with one by one input

PostPosted: Fri May 20, 2011 4:10 am
by sleepy_hollow
as what title say, is it possible for a neural network do training that provide one by one input

so this what i think, i wanna train a network with one letter input pattern, than summarize the epoch from that. after the network converged, the netwrok will provided with new input pattern, than do training again until all input are trained. the epoch are total from each epoch from each input that tranied, and output pattern will change if there are new input pattern

so is it possible that the network will recognize a character precisely as same as network which provided all input for each epoch

Re: training network with one by one input

PostPosted: Fri May 20, 2011 11:12 am
by andrew.kirillov
sleepy_hollow wrote:i wanna train a network with one letter input pattern, than summarize the epoch from that. after the network converged, the netwrok will provided with new input pattern, than do training again until all input are trained.

So on the second phase you will provide only second character to the network and the first (which was used previously) will not be provided to the network any more. Am I right? If so, the answer is NO - feed forward network with Error Back Propagation learning algorithm will not help you. You should have all the data, which is supposed to be used for training.

However, you can use 1 character on the first phase, keep it in your database and then when you get another one, you can use 2 characters to update your network (so you don’t need to teach network from random weights). However, you will need to create network of proper size/structure from the very beginning.

Re: training network with one by one input

PostPosted: Wed May 25, 2011 4:14 pm
by sleepy_hollow
However, you can use 1 character on the first phase, keep it in your database and then when you get another one, you can use 2 characters to update your network (so you don’t need to teach network from random weights). However, you will need to create network of proper size/structure from the very beginning.


is the save method does save to database?
if it doesn't, so how to save it in to a database?

Re: training network with one by one input

PostPosted: Wed May 25, 2011 8:34 pm
by andrew.kirillov
sleepy_hollow wrote:is the save method does save to database?

Save() method saves the network - nothing else. I hope it is clear from documentation.

sleepy_hollow wrote:if it doesn't, so how to save it in to a database?

It is up to you to decide what is your DB engine and how you save data. Text file, XML file, some custom file format, SQL database, whatever else you may want. This has nothing to do with neural networks and is totally application's part.

Re: training network with one by one input

PostPosted: Mon May 30, 2011 7:32 am
by sleepy_hollow
i still confused how to insert multiple binary images and then convert it into vector input as inputs to train it in a network teacher at same time

Re: training network with one by one input

PostPosted: Tue Jun 07, 2011 2:47 pm
by sleepy_hollow
since i found a way to insert all inputs, now i had a question
if i had 36 character and each character had 10 varaints, and output patterns are 36 and i use backpropagation teacher in aforge library

so, should i divide my error value?

Re: training network with one by one input

PostPosted: Tue Jun 07, 2011 3:31 pm
by andrew.kirillov
Most probably the answer to your question can be found here: RunEpoch(). The it is up to you to decide if you want to divide it or not.

Re: training network with one by one input

PostPosted: Thu Jun 16, 2011 4:01 am
by sleepy_hollow
there are another problem again

when i train my network with error doesn't divided, the error values stack at some range point

when training starts and epoch 0, the error values are 9000++, when the error values hit at 958 point and epoch are 10.000++, it will decrease 957 when epoch reach further and more further epoch it'll back to 958 point again and will decrease again when epoch reach more further

so what's wrong? i've tried to researh with alfa values

Re: training network with one by one input

PostPosted: Thu Jun 16, 2011 8:18 am
by andrew.kirillov
But what do you expect? Do you want to get 0 error? You will never get it. Do you want to get error which is always decreasing? Also hardly possible. Sometime it may slightly jump up.

Re: training network with one by one input

PostPosted: Fri Jun 17, 2011 3:09 am
by sleepy_hollow
i expect that error would decreased until reach my error limitation and be converged, but it never reach it while i use so many pair of data on it
when it just one pair data, it would reach my error limit not so long