non-linear dimension reduction via Autoencoder

hello all, I am trying to use the Matlab implementation of autoencoder to reduce the dimension of 1509 samples of Bag-of-visual word models of images, but I am surprised that while the image classification without dimension reduction recorded about 50% accuracy, and Matlab's PCA improved it to 60% but the Matlab implementation of autoencoder (with logsig activation and default values for all the parameters) reduced it to 40%. I expect higher accuracy from autoencoder, what can be the problem?

Answer 1

[BERGHOUT Tarek]

1) try to normalize you data first, between 0 and 1.

2) use these autoencoders and tell me the difference

https://www.mathworks.com/matlabcentral/fileexchange/66080-auto-encoder-based-extreme-learning-machine?s_tid=prof_contriblnk

Answer 2

[Tommi Kärkkäinen]

You might try this also:

https://se.mathworks.com/matlabcentral/fileexchange/119043-additiveautoencoder?s_tid=prof_contriblnk