WEBVTT

NOTE week11-clip1-forward
NOTE Narration read by an AI-generated voice.
NOTE The clip draws its own caption along the bottom of the frame. These cues
NOTE sit there too, so that no part of the animation is ever covered.

1
00:00:00.150 --> 00:00:04.428
Diffusion models learn to generate
by learning to undo noise.

2
00:00:04.428 --> 00:00:07.349
We start from a real backbone,

3
00:00:07.349 --> 00:00:12.549
x 0, drawn here as 16 C-alpha
points in a short helix.

4
00:00:12.549 --> 00:00:15.683
Training runs in two directions.

5
00:00:15.883 --> 00:00:20.254
The forward direction adds a little
Gaussian noise at every step.

6
00:00:20.254 --> 00:00:23.288
Each point drifts by a random amount,

7
00:00:23.288 --> 00:00:28.083
and after the first few steps
the helix is still obvious.

8
00:00:28.283 --> 00:00:33.030
You never have to take the steps
one at a time. After t steps,

9
00:00:33.030 --> 00:00:39.587
x t is a single weighted mix: the square
root of the signal weight times x zero,

10
00:00:39.587 --> 00:00:45.817
plus the square root of one minus
that weight, times one Gaussian draw.

11
00:00:46.017 --> 00:00:48.175
The schedule fixes those weights.

12
00:00:48.175 --> 00:00:55.498
With beta rising from 0.0001
to 0.02 over a thousand steps,

13
00:00:55.498 --> 00:01:00.201
the signal weight is 0.95 at step 100,

14
00:01:00.201 --> 00:01:07.483
0.72 at 250, and 0.28 at 500.

15
00:01:07.683 --> 00:01:14.887
By step 999 the signal weight
is 0.006: a Gaussian cloud.

16
00:01:14.887 --> 00:01:17.604
Every protein ends in the same place,

17
00:01:17.604 --> 00:01:22.350
which is exactly why sampling
can start from noise alone.

18
00:01:22.550 --> 00:01:27.637
Noising needs no learning at all,
just a random number generator.

19
00:01:27.637 --> 00:01:30.061
The network only ever learns the way back.

20
00:01:30.061 --> 00:01:36.480
One real structure gives a thousand
training examples, one for every step.