WEBVTT

1
00:00:00.000 --> 00:00:01.958
Dropout can copy inputs correctly...

2
00:00:02.200 --> 00:00:04.404
...and still send a gradient to the wrong place.

3
00:00:04.667 --> 00:00:08.534
At zero drop rate, each input must receive its own output gradient.

4
00:00:08.767 --> 00:00:13.170
The gradients are [1,2,3] and [10,20,30].

5
00:00:13.400 --> 00:00:16.571
The original overwrites the first and leaves the second at zero.

6
00:00:16.833 --> 00:00:18.705
The loop keeps writing to input zero.

7
00:00:18.967 --> 00:00:21.759
Using the current input index repairs the tested route.

8
00:00:22.000 --> 00:00:24.124
Finite differences confirm the derivative.

9
00:00:24.367 --> 00:00:27.981
Three added tests fail before; all 19 tests pass after.

10
00:00:28.233 --> 00:00:31.169
Full-model and Samsung-device effects were not tested.

11
00:00:31.400 --> 00:00:33.876
Read the code and evidence at gero.uz.