validation sets vs test sets

what is difference between validation and test datasets ?

Answer 1

[Sean de Wolski]

I assume you're talking about Neural Networks?

If so, validation is used for the neural network to decide when training is complete and to avoid overfitting. Testing is an independent test set.

http://www.mathworks.com/help/releases/R2015b/nnet/ug/divide-data-for-optimal-neural-network-training.html

Answer 2

[Greg Heath]

 Total     = Design + Nondesign
 Design    = Training + Validation
 Nondesign = Testing

 Total       = Training + Nontraining
 Nontraining = Validation + Testing

 Overfitting: Using more weights and biases than necessary
 Overtraining: Improving the performance of the training data at the expense of deteriorating the performance
 on nontraining data

 Training data subset: Used to DIRECTLY estimate weights and biases. Performance estimates are BIASED.

 Validation data subset: Used to 
 (1) determine when overtraining an overfit net begins to occur AND 
 (2) rank multiple designs. 
 Performance estimates are SIGNIFICANTLY LESS BIASED than training data estimates.

 Test data subset: Used to obtain UNBIASED ESTIMATES of performance on nontraining (INCLUDING UNSEEN!) data

HOPE THIS HELPS

GREG