search

TromboneDavies / PolarOps

0 stars 0 forks source link

Incorporate non-word features into classifier #59

Open divilian opened 3 years ago

divilian commented 3 years ago

Assume for now that the answer from Dave is just "yeah, stick 'em in there alongside the words, no special sauce required."

divilian commented 3 years ago

Dave says:

Yes, but in retrospect some care needs to be taken. Most NNs require the inputs to be scaled, either to have length 1 or so that all inputs are in [-1,1] or some such. You should thus scale these new features accordingly. Otherwise, the nature of the non-linear terms (the activation functions) will either make these features irrelevant or too relevant, depending on what the implementation of the NN does about this. From a theoretical standpoint, it probably doesn't make any difference. From a practical standpoint it is probably essential.

divilian commented 3 years ago

Looks good! (b02ee42) Do due diligence to make sure features are actually in there and working.

divilian commented 3 years ago

Roll these changes into sweep.py.