[Neural Network] Normalization issues when all values are the same

[joeyklee]

Dear ml5 community,

I'm submitting a new issue. Please see the details below.

→ Step 1: Describe the issue 📝

Did you find a bug? Want to suggest an idea for feature?

• Bug

Related to ml5js/ml5-website#154

It seems that if all the values of an input are the same, then the training fails. In some ways this makes sense that this would fail, but we definitely will need to handle such cases in a friendly way.

cc/ @yining1023

[hansvana]

Thanks for the reply. Might this bug also affect another weird case I've noticed? If you have a classification task and the output is a column that only holds 0 or 1, or something that normalizes to just 0 or 1, the training won't run. If you replace 0 and 1 with 'no' and 'yes' the training runs fine.

demo here

Btw: if the output column is 0 or 1 neuralNetworkData.meta.outputUnits equals 1. If the output is 'no' or 'yes' outputUnits equals 2.

If this is not related, or if you'd like me to do provide some more data let me know.

[joeyklee]

@hansvana - For classification tasks, the NN expects a string. In this case if you convert your 0 and 1 to String types, then your NN should behave as expected.

I think the first issue noted here is likely different. Thanks for bringing this up though!

[hansvana]

@joeyklee Converting 0 and 1 to Strings does indeed work. However converting to "0" and "1" raises another interesting side effect: the 0's and 1's get reversed!

I made a simple sketch to illustrate this. As you can see 0 and 1 get reversed in the output XOR feature. Replacing "0" and "1" with "False" and "True" or "no" and "yes" works as a workaround, which solves the problem for me, but I just wanted to flag the problem.

Edit: Actually after some more research the problem seems to be not just with 0 and 1. I've opened a new issue at #1044