WEBVTT

00:00:00.102 --> 00:00:01.622
The forward result is one.

00:00:02.513 --> 00:00:04.485
But the gradient is minus infinity.

00:00:05.374 --> 00:00:06.856
I reproduced this in N N

00:00:06.869 --> 00:00:09.008
trainer's C plus plus Divide layer.

00:00:10.302 --> 00:00:12.402
Both inputs are tiny and nonzero.

00:00:13.293 --> 00:00:15.535
Squaring the denominator underflows before

00:00:15.548 --> 00:00:17.648
division, even though the correct gradient

00:00:17.661 --> 00:00:18.422
is finite.

00:00:19.672 --> 00:00:22.004
Across four thousand ninety-six inputs,

00:00:22.159 --> 00:00:24.362
a local correction removed thirteen hundred

00:00:24.375 --> 00:00:25.342
two failures.

00:00:26.232 --> 00:00:28.256
Restoring the original code brought every

00:00:28.268 --> 00:00:29.158
failure back.

00:00:30.410 --> 00:00:32.806
All twenty-five focused Divide tests pass

00:00:32.819 --> 00:00:33.786
with the correction.

00:00:34.677 --> 00:00:36.610
These are CPU float tests.

00:00:37.499 --> 00:00:39.625
Effects on complete models or Samsung

00:00:39.638 --> 00:00:41.133
devices remain unmeasured.