New SSL algorithm called Maximum Manifold Capacity Representations (NeurIPS 2023)

[RylanSchaeffer]

This is a NeurIPS 2023 paper with a new SSL vision method: https://neurips.cc/virtual/2023/poster/70447

It'd be great if Lightly could add it.

[guarin]

This looks interesting, will have a closer look later :)

[RylanSchaeffer]

Sounds good!

The loss is quite straightforward. Suppose x has shape (batch size, number of views/transformations e.g., 2, 4, 32, projection output dimension). MMCR proposes the following loss:

avg_over_views = torch.mean(x, dim=1)
loss = - torch.linalg.matrix_norm(avg_over_views, ord="nuc")

[guarin]

Arxiv link: https://arxiv.org/abs/2303.03307

[johnsutor]

I'm working on implementing this now.

[guarin]

Remaining issues that we have to do before this issue is complete:

• [x] #1451
• [ ] #1452
• [x] #1481
• [ ] #1496