Application of PRIMAL-GMM

  • Synthetic Data

We use two syntheic datasets of GMMs that are generated from a linear and a nonlinear latent function to test PRIMAL with a linear HLS and a kernel HLS respectively.

Linear Synthetic Data

Nonlinear Synthetic Data

 

  • Eye Fixation Data and Flow Cytometry Data

 

Eye Fixation

 

 

 

 

 

 

 

 

 

 

Flow Cytometry

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

  • Topic Models

Visualization of word probability change in the HLS

 

  • Quantitative Result

KL reconstruction loss for held-out test GMMs and LDA classification accuracy in the latent space. The LDA accuracy is the result of the smallest KL loss among several trials. Red color denotes the best performance.