Clustering in hyperspace

[drinkingkazu]

As an extension of our past clustering attempts (dynamic-gcnn here and there, as well as Dae Hyun's recent attempt based on this), here's another architecture proposal: apply clustering loss at all spatial resolution level in U-ResNet(+PPN).

• Starting from the most spatially contracted tensor (the bottom of U), apply N convolutions, then interpret each pixel's feature array as hyperspace coordinate and apply the "clustering loss". A cluster pixel mask needs to be down-sampled w/ scn operation appropriately. For pixels where clusters overlap, I think taking the "minimum distance to any of overlapping cluster" is OK in the loss.
• N convolutions can apply >3 filters and we can still visualize (i.e. no need to be 2d or 3d) using t-SNE or something.
• The output of N convolutions should be up-sampled and concatenated to the next block in the decoding path to effectively propagate clustering information to a higher spatial resolution. It should not hurt segmentation task, but hopefully helpful if anything.

Any volunteers? :)

[dkoh0207]

Is this simply performing clustering loss as in this at the embedding space represented in the deepest layer?

[drinkingkazu]

@dkoh0207 almost but not quite. I am suggesting to do it at each stage from the bottom of U to each stage of decoding path.

[dkoh0207]

From Yesterday's Discussion:

1. Learning Embeddings

   • Discriminative Loss
   • Recurrent Pixel Embeddings
   • Spatial Embeddings

2. Breaking Translational Equivariance

   • Semi-Convolutional Operators
   • Coordconv

3. Other Architectural Improvments

   • Multi-Task Loss
   • Graph Embedding Learning

4. Applications

   • Biomedical Imaging
   • Harmonic Embeddings (Biomedical Imaging)