Neural network update via new inputs

You are here Home » Neural Network Forums » Neural Networks » General Neural Networks NN update via new inputs Primary tabs

    View(active tab)
    Edit

basic's picture published by basic on Thu, 02/27/2014 - 09:23

Hey,

I'm working on a project that is basically the design and training of a NN for the control of individual electronic components. It's fairly basic, X inputs X outputs. these components share their info (inputs) and using this shared information (I guess it's a kind of cloud computing) each component decides their optimal status. The problem is that every now and again an additional component is added to the system (X+1) and the NN needs to be updated, however since this system is supposed to work autonomously I need an easy to implement method for updating the NN without having to train it from scratch. I figured that since the system has only been slightly altered their has to be a way for this.

Thanks in advance and I appreciate any help.

Answer 1

[Greg Heath]

1. Obtain the current weights.

2. Define a new net with an extra input and output node

3. Assign the current weights to the corresponding positions in the new net

4. Either initialize the new weights or just leave them at zero.

5. Train the new net with the combined data.

Hope this helps.

Thank you for formally accepting my answer

Greg

[david]

Thanks Greg,

Have ever done anything like this? What you have written is straightforward and correct however the problem is the lack of new examples for the training. Do you know any research articles or something I can study?

[Greg Heath]

I have only updated RBF classifiers by adding hidden nodes corresponding to clusters of new data that have been misclassified.

Q: In your problem one input node and one output node are added at each structure modification? Why? What is the net supposed to be doing?