WEBVTT

NOTE week06-clip2-contrastive-learning
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 --> 00:00:05.422
Contrastive learning starts by making
two different views of the same ECG,

2
00:00:05.422 --> 00:00:08.617
using random augmentations.

3
00:00:08.817 --> 00:00:11.929
Here, view one is cropped
and has added noise,

4
00:00:11.929 --> 00:00:16.883
and view two is rescaled with a slow
baseline drift. The heart rhythm,

5
00:00:16.883 --> 00:00:20.550
and so the diagnosis, stays the same.

6
00:00:20.750 --> 00:00:24.694
Both views pass through the same
encoder, with shared weights.

7
00:00:24.694 --> 00:00:29.350
Each view becomes one point
in the embedding space.

8
00:00:29.550 --> 00:00:36.859
In practice we train on a batch, so the
other ECGs in the batch, here B, C, and D,

9
00:00:36.859 --> 00:00:39.350
also become points.

10
00:00:39.550 --> 00:00:43.091
Training pulls the two views
of ECG A toward each other,

11
00:00:43.091 --> 00:00:47.350
and pushes the other ECGs
in the batch further away.

12
00:00:47.550 --> 00:00:50.184
The loss turns this into a classification.

13
00:00:50.184 --> 00:00:56.472
Compare A1 with every other view in the
batch, apply a softmax, and its partner,

14
00:00:56.472 --> 00:01:00.683
A2, should win, here with 0.59.

15
00:01:00.883 --> 00:01:05.861
This is the InfoNCE loss. It is
a softmax over the similarities,

16
00:01:05.861 --> 00:01:09.050
each divided by a temperature tau,

17
00:01:09.050 --> 00:01:14.283
and we minimize the negative log
probability of the true partner.

18
00:01:14.483 --> 00:01:17.376
After training on many unlabeled ECGs,

19
00:01:17.376 --> 00:01:21.316
recordings with similar rhythms
tend to land close together,

20
00:01:21.316 --> 00:01:25.883
even though the model never
saw a single diagnosis.

21
00:01:26.083 --> 00:01:31.862
For pathology tiles, the two views come
from random crops, flips and rotations,

22
00:01:31.862 --> 00:01:33.819
and changes in stain color,

23
00:01:33.819 --> 00:01:39.813
because tissue has no fixed orientation
and staining varies between labs.