Scaling sensor output before training a machine learning model

Question

I'm working with the Emotiv EEG headset. My goal is to collect the signal from its 14 sensors while a subject who is wearing it performs a certain task, and use a machine learning model (a neural network in the present case) to identify certain patterns of interest in the signal from the device.

I would like to scale the sensor data before providing it as input to the model. I'm collecting this data in separate sessions for each individual subject. I observed that initially, the sensor data varies between 7,500 and 8,500. I'm considering scaling the data across all the subjects to zero-mean and unit-variance, followed by PCA or an equivalent decorrelation step. Then I'll train the model.

My questions:

(1) Would my steps for scaling the data be ok, or are there any additional precautions I need to take while dealing with the sensor data?

(2) When I'm using the trained model in real setting with a new subject, and I want it to make predictions while I collect the sensor information about the subject, how should I scale the data then? Should I scale it by the same factor with which I scaled the data used to train/test the model?

(3) And what if I would like to update my model as I'm collecting new data from other subjects in an online manner? How should I scale the sensor data then?

I will be trying different possibilities next and I was hoping to get some advice here.

Thanks!

Answer 1

(1) Your approach is correct;

(2) Use the same mean value and standard deviation of training data to scale your test data before PCA implementation;

(3) The same as (2). The general approach for the weight updates is the same between online and offline learning. The only difference is that offline learning will sum the error over all inputs (batch gradient descent), while the online learning will compute the error for each input once at a time (stochastic gradient descent). But both should use the same scaling method.