5 ms·
We blow up model sizes to reduce the risk of overfitting and to speed up training. So yes, usually you can shrink the finished model by 99% with a bit of normal
by fxtentacle 4y ago
We blow up model sizes to reduce the risk of overfitting and to speed up training. So yes, usually you can shrink the finished model by 99% with a bit of normalization, quantization and sparseness.
Also, plenty of "deep learning" tasks work equally well with decision trees if you use the right feature extractors.
- jakearmitage 4y agoWhat are feature extractors?
- danuker 4y agoI suspect features created manually from the data (as opposed to solely using the raw data): https://en.wikipedia.org/wiki/Feature_(computer_vision)#Extraction https://en.wikipedia.org/wiki/Feature_(computer_vision)#Extr...
- optimalsolver 4y agoI thought increasing model size raises the risk of overfitting, rather than the opposite?