I would like to use VAE model in unsupervised learning to generate new feature. Most of the examples are supervised and semi-supervised learning. Where can I find for unsupervised learning or can it be possible in unsupervised learning for generative model?


I use my training data ( 1,42 ) features with no label. Most of the example for VAE (i.e MNIST data) used images. But in my case, I only use 2D coordinates to train VAE. My data are (x1,y1,x2,y2,...., x42,y42), (1,42) dimension.

Sample data: `[297.425   341.30002   280.1   295.625 275.375 240.5   287.975 213.725 294.275 186.95  332.07498   254.675 355.69998   215.3   380.9   201.125 402.94998   188.52501   357.275 268.85  391.925 234.20001   412.4   215.3   432.875 202.7   380.9   287.75  410.82498   259.4   432.875 238.925 450.2   224.75  391.925 306.65  428.15  290.9   448.625 272 469.1   254.675]`

And train and predict. But, when I predict, my reconstruction is not good even I get good training accuracy. I want to get back true or nearly x1,y1,x2,y2,...,x42,y42 reconstruction coordinates. That is why I used VAE model to generate new features based on training features. I will use this in noisy reconstruction.


closed as unclear what you're asking by mkt, Peter Flom Mar 28 at 11:10

Please clarify your specific problem or add additional details to highlight exactly what you need. As it's currently written, it’s hard to tell exactly what you're asking. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.


Variational autoencoders are unsupervised learning methods in the sense that they don't require labels in addition to the data inputs. All that is required for VAE is to define an appropriate likelihood function for your data. Commonly, a reconstruction loss is used, e.g. binary cross-entropy, squared distance, or absolute distance, to compare the actual input to the VAE with the generated output of the VAE. If you want to use VAE for dimensionality reduction or feature extraction, you would feed our data to the encoder network and use its output for downstream tasks.

Keep in mind that VAE is a generative model, thus training encourages the output of the encoder to resemble an isotropic Gaussian distribution. If you don't need to generate data and just want to extract features, you might get better results with a traditional autoencoder without the Kullback-Leibler divergence regularizer that is used in VAE.

  • $\begingroup$ Yes, I know VAE is a generative model. That is why I want to use this model for my data. My data have (1,42) coordinates, I want to get back true reconstruction after prediction vae model. But, when I reconstruct, the result is not good. And I only train coordinate values without image pixel (e.g. mnist data have (28,28)) or anything. $\endgroup$ – Dennis Thor Mar 28 at 7:39
  • 2
    $\begingroup$ @DennisThor so what exactly is the problem? You asked if VAE can be used in unsupervised scenario, and the (correct) answer is: yes, they can because it is an unsupervised learning algorithm. But in comments you say something about achieving poor results when using VAE, this seems completely unrelated to the question you asked? $\endgroup$ – Tim Mar 28 at 10:48

Not the answer you're looking for? Browse other questions tagged or ask your own question.