I have been using PLSA to reduce the dimension of image BOV models due to its ability to capture the co-occurrence of visual-words. However, the linearity of PLSA limits its accuracy when applied to the classification of image collections with high and medium complexity. I have recently discovered that Deep Learning algorithms provides non-linear approaches for machine learning tasks. which of the deep learning approaches is the most suitable for dimension reduction?
I think the most popular one is deep autoencoders
Useful link
https://deepvision.forge.nicta.com.au/2014/GAE_DeepVision.pdf
A comprehensive tutorial on Deep Learning by J.Bengio
http://www.iro.umontreal.ca/~bengioy/talks/aaai2013-tutorial.pdf
You can find a step by step tutorial for deep learning in this link, http://deeplearning.stanford.edu/tutorial/ also you can read more about it in, deep learning.org
Maybe the deep belief network (DBN) is a good choice. You can find some new paper using DBN-based model for dimension reduction.
Article Reducing the Dimensionality of Data with Neural Networks
- Badges
- Science topic