callowbird / harmonica Goto Github PK

In the subsection, under the title A.2 importance features, of the appendix, the article mentioned that based on Table 3, Harmonica concluded that the Initial learning rate of the small network and for the large network is in the range from 0.001 to 0.1. (At stage 1-3, 04. Initial learning rate *05. Initial learning rate (Detail 1))

My question is: how can we conclude this statement from the 4th, 5th and 6th options (Initial learning rate) ? For example, If "-1" stands for "T", (x_4, x_5, x_6)= (-1, -1, -1) means the initial learning rate=0.3. Do I take this right?

If I do, then since Table 3 suggests x_4* x_5 is important, I might get one of the ranges, >= 0.1, [0.01, 0.1], [0.001, 0.01], or <= 0.001.

The paper seemed to locate none of them.

If you wouldn't mind, please add a liberal license file named LICENSE.txt, having for example a BSD or MIT license. You can find the MIT License at https://choosealicense.com/licenses/mit/

Could you give some guide lines of the whole section? For instance, is the sparse vector "s_i,j" in this section the same as the one in the proof of Theorem 6 which is a "s-sparse vector x" ? And why does it contain 5 pairs of weight and feature?