PRLab
TUDelft
NL
What is the source of this magic?
 Deep net has more parameters than the shallow net;
 Deep net is deeper than the shallow net;
 Nets without convolution can’t learn what nets with
convolution can learn;
 Current learning algorithms and regularization
procedures work better with deep architectures than
with shallow architectures;
 All or some of the above;
 None of the above?
PRLab
TUDelft
NL
Do Deep Nets Really Need to be Deep
 Lei Jimmy Ba
University of Toronto
 Rich Caruana
Microsoft Research
 arXiv:1312.6184
PRLab
TUDelft
NL