Feature weighting

I am currently training an architecture for regression purposes.
The training data consists of images and their ground truth labels (which is a 15x1) vector representing information relevant to my purpose.
Out of these 15 features,some are important while others can be ignored.

As such is there a way to assign weight to these features and then train the network so that the final prediction is consistent with the weights assigned during training.
I read somewhere that feature weighting shouldn’t be done


My approach:
let’s say I have about 2000 samples ,so my training data will be of size 2000x199
I can compute the covariance matrix of this and use that to determine the weights for the features.
Would that be correct?