feature vector of NN

how to find the feature vector of NN in the attached results obtained from the nprtool (which is app for develop pattern recognition appalication of Neural network).

[Athul Prakash]

I'm not aware of a single function or a direct way to extract intermediate layer outputs, but I have suggested a couple of workarounds in my answer below.

[Yogini Prabhu]

thanks i will try those

Answer 1

[Athul Prakash]

You can extract the network and inputs from the 'results' struct using..

net = results.net
inputs = results.inputs

I found 2 ways to extract activations from intermediate layers in a 'network' object :-

Approach 1

You may modify the architecture to specify the required layer as an outputLayer..

% Say 'net' is the network. Assume you want the output of layer 'i'.
net.outputConnect(i) = 1 % Modifies the architecture to set layer 'i' as an additional output.
out = sim(net, inputs) % the activations of i-th layer are now included in the output vector.

Approach 2

Manually run the neural network by executing it's underlying maths.

Suppose you're trying to get the outputs of layer 3...

X = inputs; % Initialize 'X' as inputs. Each layer's values will be computer into 'X'. 

% Layer 1
X = net.IW{1,1} * X + net.b{1}; % Compute the values of layer 1. Use 'net.IW' (input weights)
X = tansig(X) % apply whichever is the activation function of layer 1.

% Layer 2
X = net.LW{2,1} * X + net.b{2}; % compute the values of layer 2. Use 'net.LW' (layer weights)
X = logsig(X) % apply whichever is the activation function of layer 2.

% layer 3
X = net.LW{3,2} * X + net.b{3} % same of for layer 2
X = softmax(X) % apply whichever is the activation function of layer 3.

% 'X' now contains the activations of the third layer in the network.

The manual computation might require tinkering with the 'network' object's many properties. You may follow the attached documentation to get a better understanding of these properties and how to manipulate them.

https://www.mathworks.com/help/deeplearning/ug/create-and-train-custom-neural-network-architectures.html

Hope it Helps!