Menu

Manifold sculpting seems to produce wrong results.

Help
Anonymous
2015-09-08
2015-09-08
  • Anonymous

    Anonymous - 2015-09-08

    I want to implement the example of letter 'A' on the wiki page of nonlinear dimensionality reduction(https://en.wikipedia.org/wiki/Nonlinear_dimensionality_reduction). I use python to generate the data, and then use waffles to do dimensionality reduction. However, the result of PCA is a cicle while the result of manifold sculpting is some disordered points. Could someone help me with this problem?

     
  • Mike Gashler

    Mike Gashler - 2015-09-08

    I did that a long time ago, and have since forgotten some of the details. However I found this in a search of my computer: http://uaf46365.ddns.uark.edu/letter.zip
    It may or may not be the data that produced that figure. The file log.txt appears to contain some notes about the exact commands I used to produce varying quality of results. Apparently, it took some experimentation before I was happy with the results. I am not sure whether the best one is in there or not.

    (Tangentially, if you're looking for an NLDR algorithm that just works in a variety of cases, frankly, I'd recommend looking into some of the neural network-based feature-learning methods. Manifold Sculpting can produce good results in specific cases, but it requires careful tuning and a lot of processing time. Ultimately, all of the NLDR methods that require neighbor-finding as their first step end up being too slow for general applications, too susceptible to problems with poorly-sampled manifolds, and unable to generalize effectively for out-of-band samples. The neural network approaches, by contrast, require no neighbor-finding step, and are designed to generalize.)

     

Anonymous
Anonymous

Add attachments
Cancel





Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.