Neural Network ToolBox : Proper function to train multilabel data (Backpropogation )

Hello,i m a matlab beginner..

Which Inbuilt functions are suitable to train multilabel dataset ? using backpropogation ?

i also want to :

1.take outputs of that function at output layer and then modify it (while training) after each epochs

2.after getting output , define our custom error measures like hamming loss ,ranking loss etc (specifically for multilabel classification)

 i know how to modify it , but is it feasible to do all these things with inbuilt function?

if yes,then which function should be used here ? and where i can get its tutorials ?

Thank you for your consideration... Please Help !!

[Greg Heath]

Unfamiliar with the terms multilabel, hamming loss and ranking loss. Apparently you don't mean multi-variable or multi-class.

Please explain the difference.

Searching multilabel:

NEWSGROUP search 0 hits

ANSWERS search 1 hit (besides your 2)

MATHWORKS.COM WEBSITE SEARCH 4 additional hits (2 file-exchange)

When I get a chance I will look at these.

Meanwhile, will you explain a little more?

Greg

[Greg Heath]

OK I think multi-label is just non-mutually exclusive classes. See my answer to your other post.

 help newpr
 doc newpr

 help patternnet
 doc patternnet

Watch out for possible mutually exclusive class assumptions.

What modifications do you want to do after each epoch?

I'm a little concerned about your intentions to modify the learning process.

Not familiar with your loss measures. Can you explain?

[Greg Heath]

Wikipedia explains hamming loss. What about ranking loss?

[pooja]

hello sir multilabel is also called multi class multi label

that means out of different classes , an instance may belong to more than 1 class labels

for example

classes : comedy ,horror ,suspense, romentic ,action

a movie can be comedy as well as action

then an instance have classes [1,0,0,0,1]

means as an output i may get more than 1 class-labels on ith instance

i hope i explained my point well

[pooja]

multilabel data ex: i have 3 instances here [ 1 1 1 2 , 1 2 3 1, 4 5 2 1] all having 4

features/attr in it and each instance belongs to [0 1 1 , 1 1 0, 0 0 0]

0,1 or more than one classes out of three

ranking loss : absent labels in instance are ranked higher than present labels in it

[pooja]

Your question :*What modifications do you want to do after each epoch?*

"after each training epochs , i want to access output layer and want to convert all the values in its corrosponding log values ,it will help in updating weights and bias values in hidden layer"

[Greg Heath]

You are wrong in trying to manipulate after each epoch. It just makes training take longer. Perhaps using the entropy fnction for non-mutually exclusive classes will help.

Answer 1

[Greg Heath]

>Which Inbuilt functions are suitable to train multilabel dataset ? using backpropogation ?

 Multilabel is just classification with non-exclusive classes. 

 Use patternnet with targets in {0,1}

 HOWEVER, the relation between target and class indices is NO LONGER given by vec2ind and ind2vec.

>i also want to :

>1.take outputs of that function at output layer and then modify it (while training) after each epochs

You will have to train in a loop over 1 epoch design stages. Training time could be a problem.

>2.after getting output , define our custom error measures like hamming loss ,ranking loss etc (specifically for multilabel classification)

 Wikipedia defines hamming loss. Never heard of ranking loss. How do you define it?

>i know how to modify it , but is it feasible to do all these things with inbuilt function? if yes,then which function should be used here ? and where i can get its tutorials ?

help patternnet
doc patternnet

Hope this helps.

Thank you for formally accepting my answer

Greg