CNN not working after standardizing data

Deep Learning
1

I ran into a very weird issue with a CNN I was training.

My CNN loss was not changing, and upon looking at the output predictions of the softmax, I saw that on all samples, all the predictions were approximately the same.

The CNN had multiple input channels, so for bug testing purposes, I just only started using 1 channel, but the issue persisted.

For my data, I had been standardizing (zero mean, unit variance) my data. So I then tried stopping the unit variance part of the standardization, and then the CNN started learning again. But this is really weird, since everyone generally standardizes their data. Since without making them unit variance, my input data values can be from -300. to +300., which seems like it would be hard for learning for the CNN, but it is somehow doing fine with such a large range.

Does anyone have any insight into why this is happening?
I ultimately do want to standardize my data since I have different sensor data and they will have different ranges.

2

I’m having the same problem. Using a CNN with Keras. With standardized data I keep getting the same values for prediction, even when changing model size. Without standardization it works kind of okay.

Have you found a solution?