WEBVTT

00:00:00.000 --> 00:00:06.500
At xB, the reverse-KL A term is minus 0.243279 and the B term is plus 0.554518.

00:00:06.500 --> 00:00:13.500
The xB total is 0.311239 nats; the x total is 0.130812 nats; averaging full-precision totals gives 0.221025 nats.

00:00:13.500 --> 00:00:20.000
Only after the loss exists do we update student parameters. No gradient passes through the discrete draw, and no post-update probabilities are asserted.