assuming you are correct, DL is simply data fitting as you claimed. data fitting is not only about parameters and its number, it is also about the best group of functions to be fitted.
all the different nn architectures: cnn, rnn, lstm,are about what group of functions to pick to fit. These architectures were developed before 2000. We did not have the computation power to test out these arthitectures with large amount of data until 2012.
if old architectures have no use, you can develop a "shit" new fundamentally different nn architecture and fit it with billions parameters to see whether it works or not.
fitting is not only about parameters and its number, it is also about the
best group of functions to be fitted.
all the different nn architectures: cnn, rnn, lstm,are about what group of functions to pick to fit. These architectures were developed before 2000. We did not have the computation power to test out these arthitectures with large
amount of data until 2012.
if old architectures have no use, you can develop a "shit" new fundamentally different nn architecture and fit it
with billions parameters to see whether it works or not.
这篇有没有道理
如果这种都看，世上的paper是看不过来的。
数学家和物理学家都觉得他们可以来掺一脚，
其实不是那样的。CS有CS的套路。
看paper主要是看evaluation。
CS要成为Science，这才是正确的路数。
数学家和物理学家都觉得他们可以来掺一脚，其实不是那样的。CS有CS的套路。
谁签字了？
人脑就是量子计算
【 在 neversee (山知了) 的大作中提到: 】
: 想想生物基因大数据，超级计算机都还吃力，大自然一点不费劲，分明就是量子计算。
: Gerard 't Hooft说宇宙就是量子计算机！CS搞的一套在上帝面前就不好意思了。