Why neural network gives negative output ?

I have 15000 dataset, 6 inputs and 12 outputs. Using feedforward net, I get training, validation, test and over all regression above 95%.

But when I check trained net with new inputs, I get negative values in the outputs.

(There is no negative values in the dataset)

What is the reason for it?

What could be the worng?

What should I do to overcome this issue?

Answer 1

[Greg Heath]

How different is the new data (e.g., Mahalanobis distance)?

If you know the true outputs, how do the error rates compare?

If you want positive outputs, use a sigmoid in the output layer.

Hope this helps.

*Thank you for formally accepting my answer*

Greg

[Harsha M V]

@Greg

The new input data is in within the same range of dataset.

For few inputs, I get good accuracy [near 100%], but for someother inputs the error is too high.

[Greg Heath]

Have you compared those inputs?

[Harsha M V]

Yes, the mahal distance is 6.5

[Greg Heath]

It is not uncommon for new data to lie outside the bounds of training data.

Take into account whether negative values have meaning.

If not, use sigmoids in the output layer.

Greg