artificial neural network question

Hi all,

I have scaled the input and target data by using these commands [pn,ps] = mapminmax(xt1'); [tn, ts] = mapminmax(yt1'); Should I scale the tested data as well? If yes should I use the same command?

Thanks Aziz

Answer 1

[Greg Heath]

For most of the training algorithms, scaling is an automatic default. Which algorithm are you using? Classification/pattern-recognition or regression/curve-fitting?

Hope this helps.

Thank you for formally accepting my answer

Greg

[Abdulaziz Abutunis]

Thank you Greg for the concern. And I have read a lot of answers written by you, and the were really helpful. ANN is for curve fitting and then data prediction. I am using newff command. I have scaled the tested data and reversed it after conversion achieved and the results were good. Another question emerged which how can I assign and divide the input data to trained,validated and tested data. And how I keep the process iterating till the validation output meet the satisfactory criteria.

[Greg Heath]

1. You don't have to scale the data. Normalization of inputs and targets followed by denormalization of the outputs is an automatic default.

2. I normalize the val and test data with the parameters of the trn data. I'm not sure how the NNToolbox does it ... maybe using all of the data?

3. Random data division is an automatic default (dividerand). It can be replaced by other types (search divideind, divideint, divideblock and dividetrain)

4. Validation stopping is an automatic default, provided you have not defined the validation subset to be empty.

[Abdulaziz Abutunis]

Thanks again Greg

Answer 2

[m Whelan]

An artificial neural network was trained to obtain a face recognition system of various people faces. Images of 10 people were used including 40 images per person. Each image of the database has the size of 24 x 30 pixel. The input to the network are pixel intensity values ranging from 0 to 255 which were scaled to range from 0 to 1. The network has one layer with 20 hidden units and each output unit in the network represents one of the 10 persons to identify. The image dataset was divided into 200 images for training, 100 for validation and 100 for testing.

How many units does the network have in total? Note that the network structure is a layered network with input units, hidden units and output units. Indicate one way to simplify the structure of the network.