RProp MLP Learner weight update

New Knime user. In the RProp MLP Learner Node for neural network training, how does the node know what Error function is being used? i.e. in order to determine the sign - and thus direction - of the weight changes, the partial derivatives of the Error function wrt each weight must be calculated. Where is it told what Error function to use in a specific use case/workflow?

Hi @LaurenceB and welcome to the forum.

I went digging into the code a little bit but the particular error function didn’t jump out at me. Let me ask one of our developers and see what I can find out.

Original Rprop paper is here for those that might be curious, but it deals with arbitrary error functions: https://paginas.fe.up.pt/~ee02162/dissertacao/RPROP%20paper.pdf

After checking, the error function for this particular node should just be MSE.

Note that this node is a old implementation of MLP - you could use KNIME’s newer DL4J or (better yet) Keras integrations for additional customization options in setting up a neural network.

Thanks Scott. Got it. IN general, are we able to see the exact algorithmic/mathematical/data processing choices used in the nodes which implement the algorithms? Are we able to see the code? Curious… Thanks again! Laurence

You can dig into the AP open source code here: https://github.com/knime