instead of finding ANN for the map X-->Y Try to find ANN for the map X*(1-p)+Y*(p) --> Y for many p where p is not 1 Then by those many ANN(p), using interpolation or other methods to find the ANN(1)
【在 g****t 的大作中提到】 : instead of finding ANN for the map X-->Y : Try to find ANN for the map X*(1-p)+Y*(p) --> Y : for many p where p is not 1 : Then by those many ANN(p), using interpolation or other methods to : find the ANN(1)
c*3
28 楼
我现在咋觉得变成了类似量子力学vs相对论的节奏
g*t
29 楼
(1) Possible implementation: Train Y+p*(X-Y) --> Y numerically for random samples value p1,p2,...with N(1,s). The mean of the resulted ANN should be OK. The methodology is very fruitful in history since Newton. One of the name of this approach is "continuation perturbation". I think the Residual Network belongs to this tradition. (2) Not only p,1-p average, we can use other paths to connect f1:[the unknown_ANN] and the f2:[identity map].
【在 w***g 的大作中提到】 : 神操作。可行性存疑。
h*c
30 楼
instead of finding ANN for the map X-->Y Try to find ANN for the map X*(1-p)+Y*(p) --> Y p=0.5 0.5X + 0,5Y -> y X -> y p=0.3 0.7x +0.3 Y -> y still X -> Y looks like theologically circular?
c*v
31 楼
try x = 1 and y =2
【在 h**********c 的大作中提到】 : instead of finding ANN for the map X-->Y : Try to find ANN for the map X*(1-p)+Y*(p) --> Y : p=0.5 : 0.5X + 0,5Y -> y : X -> y : p=0.3 : 0.7x +0.3 Y -> y : still : X -> Y : looks like theologically circular?