为何要“深”？
pluskid的博客 Deep Learning and Shallow Learning
Bengio Y. Learning deep architectures for AI. Foundations and trends® in Machine Learning, 2009
Deeper is Better?
模型有更多的参数会有更好的结果，这是毋庸置疑的。
深瘦的模型会比浅胖的模型有更好的表达能力。
Universality Theorem
虽然理论上单层网络可以表达任意的函数，但是实际上更深的结构在表达函数的能力更出色。
细节见 A visual proof that neural nets can compute any function
Do Deep Nets Really Need To Be Deep? (by Rich Caruana)
更多细节见 Rich Caruana
“Do Deep Nets Really Need to be Deep?”阅读笔记
参考文献
Home: http://speech.ee.ntu.edu.tw/~tlkagk/courses_ML17.html
A visual proof that neural nets can compute any function
Rich Caruana
Deep Learning: Theoretical Motivations (Yoshua Bengio)
Connections between physics and deep learning
Why Deep Learning Works: Perspectives from Theoretical
Chemistry
