How to set weights of data sample for neural net training

Hello, I am trying to train the simple shallow neural net.

However, individual data records (samples / input vectors / instances) have different importance (weight), in my task.

Traditional code (from Matlab example) for neural network training is something like this :

x = [0 1 2 3 4 5 6 7 8];  % input attribute
t = [0 0.84 0.91 0.14 -0.77 -0.96 -0.28 0.66 0.99]; % target attribute

net = feedforwardnet(10);
net = train(net,x,t);
y2 = net(x)

but I want to set also weights (importance) of individual instances, like this

w = [1 1 1 1 1 0.2 1 1 4 0.1]; 

(most of instances have normal importance (value 1), 0.1 and 0.2 are less important, and 4 is very important)

And I want to use this weights w in training proces.

It should work (theoreticaly) by multiplying the learning rate coefficient by the weight of the actual instance.

LearningRate for specific instance = network learning rate * instance weight ;

In Weka (programmed in Java), It could be set with method Instance.setWeight(), like this :

     for (int i = 0; i < data.numInstances(); i++) {
       data.instance(i).setWeight(w[i]);
     }

and, in next training process, these weights are used automatically.

Could you give me some hint, or inspiration, how to achieve this goal in Matlab (training with instance weights) ?

What are you suggest ?

Of course, I could call the Weka library from Matlab. However, Weka contain only limited types of neural networks.

Thank you.

Answer 1

[Srivardhan Gadila]

Refer to Weight and Bias Values to know on how to acess the weights & bias values and also understand the structure.

Then refer to Create, Configure, and Initialize Multilayer Shallow Neural Networks & configure and initialize the network first.

net = configure(net,x,t);

Then set the custom weights as follows:

net.LW{2,1}         % Accessing the existing weights
CW = 100*ones(1,10) % Replace this with your weights
net.LW{2,1} = CW

Then you can train the network with the updated weights. In addtion refer to Neural Network Object Properties & Neural Network Subobject Properties to understand more about the different properties of shallow nerual networks in MATLAB.