I'm super new to Neural Nets and machine learning in general, so bear with me. I guess my question is "How capable are neural nets?" I've worked with the example of training a net to predict if an 8x8 image is a "4" or a "3". Currently, I am feeding in each value into my neural net:
[ 1,0,0,1,0,0,0,0, 1,0,0,1,0,0,0,0, 1,0,0,1,0,0,0,0, 1,1,1,1,1,1,0,0, 0,0,0,1,0,0,0,0, 0,0,0,1,0,0,0,0, 0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0 ],
Kind of hard to see but there is a '4' in there. I have successfully trained a my neural net to recognize 4's vs 3's pretty effectively. But now I am working with some acceleration data to predict different motions.
I've read that a multiple hidden layers allow you to emulate non-linear functions. But I don't remember enough about linear algebra (not to mention non-linear algebra) for that to make much sense to me. , but then I started to hear about feature selection, and basically creating summary statistics about my data to use as inputs rather than raw values. I understand the need to do that when you get into larger data sets where you would have an 80x80 matrix instead of an 8x8. But is there a reason to do feature selection on smaller data sets in order to give the net more insights about the data? Or does the net able to take into consideration basically all of the nuances of the data?