|
Hi all, Is neural network is suitable for unsupervised learning like clustering. In case of its possibility, can someone refer me some related resources? |
|
Neural networks are often used for feature extraction and dimensionality reduction. Google autoencoders and RBMs. One of the original papers is this one. For clustering, I guess you would be looking at SOMs and Neural Gas. These are big in the neural network community but I have the feeling they are a bit frowned upon in the machine learning community. My bet would rather be to extract features or learn a metric using an autoencoder (my current favourit is this) and then use k-means.
This answer is marked "community wiki".
SOMs are generally considered more for visualisation than inference, but are very useful tools.
(Feb 11 '12 at 18:17)
Robert Layton
|
|
Andreas, I have high dimensional data (~ 1000 features and 1050 data points). I want to reduce the dimension, cluster and visualize them in 2-D. what is good approach in deep-learning context for this type of problems ? I've found that t-SNE can produce good visualizations, even when the data are high dimensional. This page has code and samples: http://homepage.tudelft.nl/19j49/t-SNE.html
(Aug 26 '13 at 18:12)
Daniel Hammack
Thanks Daniel, but I used R-implementation of this method for my data set, it doesn't create any plot which you can identify your clusters. but in the paper they have plots which clusters are clear with different colors. do you also know other deep learning methods for my scenario ? (clustering-> dim-reduction-> visualization)
(Aug 26 '13 at 18:51)
Azim
|
|
Neural gas based and SOM based algorithms work in unsupervised manner. Same for rbm, but i don't know if it is good without supervised fine-tuning. You could try this new neural gas based algorithm for unsupervised learning http://haselab.info/soinn-e.html |
