Open cfoster0 opened 3 years ago
An alternative to the symmetric cross entropy loss, we could optimize for alignment and uniformity on the unit hypersphere as was done in this paper:
https://arxiv.org/abs/2005.10242
The code itself is like 4 lines of PyTorch, with the caveat that we'd need to write a pdist JAX function in order to use it.
pdist
May be worthwhile to play with an additional speech-domain loss in the form of DINO, which should be relatively easy to implement, with a bit of JAX code and some crop augmentations.
https://arxiv.org/abs/2104.14294
An alternative to the symmetric cross entropy loss, we could optimize for alignment and uniformity on the unit hypersphere as was done in this paper:
https://arxiv.org/abs/2005.10242
The code itself is like 4 lines of PyTorch, with the caveat that we'd need to write a
pdist
JAX function in order to use it.