I'm using a CNN to model a problem that involves precise numerical values from a physical simulation.

After months of design/redesign and optimization, I've noticed that the majority of the "error" in this network is accumulated because of lower accuracy for smaller valued cases, whereas it seems to work fine when the values are numerically larger.

The input values are always normalized to $(0,1)$ when fed into the network.

The error is measured on the denormalized output values.

How can I make sure that the network learns these "small" values better?

I believe that the key problem lies in the fact that the network is "blinded" to the actual size of the I/O pairs because of the normalization process.

Is this an inherent problem with the normalization process, or is this a problem in how the network is designed?

The error is quantified using MAPE, in this case.


1 Answer 1


It seems you are working on a regression problem with a strictly positive target and use MAPE as the loss function to train the neural network. This loss function could be the culprit. As illustrated nicely in this answer, minimizing MAPE leads to predictions that are biased towards smaller values. This is a consequence of the asymmetry between errors in the positive and negative directions (negative errors are 100% at worst, while positive errors are unbounded).

A possible alternative is to minimize the mean absolute error computed with the logarithm of the target. It will converge to the logarithm of the target's median.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.