WEBVTT

1
00:00:00.000 --> 00:00:05.400
Zero gradients before the next optimization step.

2
00:00:05.400 --> 00:00:10.800
The forward pass turns a batch into predictions.

3
00:00:10.800 --> 00:00:16.200
Loss measures the gap between predictions and labels.

4
00:00:16.200 --> 00:00:21.600
Backward computes gradients for each trainable parameter.

5
00:00:21.600 --> 00:00:27.000
Step updates parameters; a lower held-out loss is evidence, not a guarantee.