Does the inputs and the neurons in a single hidden layer have any relation with the multiclass single output. For example, I have labels for a multiclass output ranging from 1-9 along with 3 input variables. If I choose #neurons as the mean of input and output ie (3+1)/2=2, then my accuracy for class no 7,8,9 is always 0. Why so? Is there any restrictions to the max no of classes that can be used for a multiclass classification problem?
Please suggests some solutions so that I am able to classify all 9 classes separately.
Thanks in advance.
Hi Manas, i didn't understand at all the question you asked. The number of patterns that a neural network can clasify is restricted to the number of neurons in the output layer. Also the number of neurons in hidden layer determine the complexity of patterns that the network will classify. check that your training set is adecuate for training all the possible cases.
Hi Manas,
I believe that the number of neurons in the hidden layer is tuned empirically. THis is done by trying your prediction framework using different number of neurons in the hidden layer and testing it on your testing data (preferably with cross validation). The number of neurons that achieves the highest accuracy would be selected for prediction. There is no much logic behind the number of neurons as it is the case with most of the parameters of the neural networks.
HOpe this helps.
The middle layer is called a hidden layer, since the neurons in this layer are neither inputs nor outputs. The term "hidden" perhaps sounds a little mysterious - the first time I heard the term I thought it must have some deep philosophical or mathematical significance - but it really means nothing more than "not an input or an output"
I agree with Mehiar Dabbagh,
I would only add that if you use a lot of neurons the network could suffer of overfitting so the tuning is a very difficult part of this kind of job. validation and test set could help your choice.
mariano
I would try to make it more clear. I have a problem statement where I have to classify an event out of 9 events based on 3 input variables. Change in a particular set of input may lead to any one of the 9 events.
Here the problem which I face is, even if I have ample amount of training data with training accuracy of around 90%, the testing accuracy is very less. The testing data being a subset of the training data, the events 7, 8 ,9 is never identified.
The training data size is 2000 X 7 with each class having around 230 X 7.
230 * 9 classes = around 2000 X 7
Again the same data is tested class by class. Firstly all data from class 1, then from class 2 and so on till class 9. the accuracy for a particular training is around 65% for class 1-6 and 0% for class 7-9.
Hope I was able to explain this time.
Please suggest some solution.
Hi Manas,
First I think that it is "imbalance data set" problem. However from your explanation (" 230 * 9 classes = around 2000 X 7"), this is not the case. I suppose that the training set is balanced.
My next question is about the accuracy you mentioned ("the accuracy for a particular training is around 65% for class 1-6 and 0% for class 7-9"). Do you mean this is the accuracy tested by known sample (not unknown sample)? If the answer is yes, this is not overfitting problem. On the other hand, it may be underfitting problem. How many epoches have you trained your ANN? Is it set to be more than 2000 epoches? If the answer is yes, then the only explanation I can give is your 3 input variables are not able to distinguish class 7-9.
My last question, Do you have 9 output nodes in your output layer?
Dear Manas Ranjan Prusty,
Could you please say how many hidden neurons you used?
- Badges
- Science topic