martin_a_vinck - Network

Very happy that this new paper led by PhD student in our lab Jinke Liu is now finally out: #PLOSCompBio: Improved visualization of...

Very happy that this new paper led by PhD student in our lab Jinke Liu is now finally out:

#PLOSCompBio: Improved visualization of high-dimensional data using the distance-of-distance transformation

Jinke shows here how in high-dimensional spaces tSNE and UMAP have a crowding problem due to scatter noise, which invades the cluster regions.

This problem can be solved by using a distance-of-distance transformations, where distance is now defined by similarity of distances to the local neighbourhoods of two points.

Attached an example of how this improves the tSNE embedding.

Jinke shows that this technique leads to improved visualization of high-dimensional neural representations during different network states, as well as image representations in convolutional neural networks.

https://dx.plos.org/10.1371/journal.pcbi.1010764