- Working on the gram matrix. Put the pair-wise distance to the robust auto-encoder and comparing the result with normal auto-encoder. If some features are unexpected, it should have unusual element-wise outlier on gram matrix. The robust auto-encoder should be able to filter these element-wise outlier to "S". In this experiment, my pre-experiment idea is based on the gram matrix contain redundant information thus could be compressed by the robust / normal auto-encoder.
- Add/make the W being shared variable to save it or to make prediction by adding fine tune phase.
{{ message }}