Neural Network: Custom Loss Function: Minimize Range/Amplitude

Hello,

I have a dataset of errors and an LSTM neural network which is predicting these errors.

Overall, the network is doing a decent job, if I substract the predicted error from actual error on my test-set, the error is reduced by about 80%.

However, my max(abs(error)) (error = target - prediction) stays about the same, because the network overcorrects sometimes too much, which is a highly undesirable property for my case.

To handle these overcorrections, I would like to change my loss function.

Is there a possibility to tell the network to not minimize MeanSquaredError (as the last regressionLayer does it), but to tell it to minimize max(abs(error)) ?, i.e. I would allow the mean-error to be higher if the overall corrected maximum error will be lower

Thank you for any help

Answer 1

[Abolfazl Chaman Motlagh]

From 2021a version, Matlab allows you to define and use customize data, layer, training loop, network and in as you need loss function. for more information follow below link:

https://www.mathworks.com/help/deeplearning/ug/define-custom-training-loops-loss-functions-and-networks.html

for example for how to defien you own custom network or training, these are good examples:

• https://www.mathworks.com/help/deeplearning/ug/train-generative-adversarial-network.html
• https://www.mathworks.com/help/deeplearning/ug/train-network-using-model-function.html
• https://www.mathworks.com/help/deeplearning/ug/train-a-siamese-network-to-compare-images.html

but generally it made your work hard to follow all these from scratch.

so maybe just create a simple custom layer for your output and train your network. see this documentation:

• https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html
• https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-output-layers.html