I have used a stacked denoising autoencoder to reduce dimension of a feature matrix from 18930 to 60 and the values of the original matrix are mostly zero and others are less than 0.001.
The autoencoder has 3 layers in encoder part and 3 layers in decoder.
The problem is, the encoded matrix has several columns with value of zero and other columns have nearly same values, also the loss value converge after few epochs.

This is a part of the encoded matrix:

array([[0.        , 0.        , 0.67260367, ..., 1.8776112 , 1.2193491 ,
        1.3573663 ],
       [0.        , 0.        , 0.68382835, ..., 1.9114548 , 1.2421154 ,
        1.3819823 ],
       [0.        , 0.        , 0.67055804, ..., 1.8722168 , 1.215465  ,
        1.3533425 ],
       ...,
       [0.        , 0.        , 0.66787165, ..., 1.8641704 , 1.2104    ,
        1.347388  ],
       [0.        , 0.        , 0.6682701 , ..., 1.8654424 , 1.2113018 ,
        1.3486449 ],
       [0.        , 0.        , 0.6680883 , ..., 1.8648231 , 1.2112418 ,
        1.3477939 ]], dtype=float32)

I changed learning rate, noise rate and batch size but weren’t effective.