Is there a way to introduce sparsity Fourier/Cosine/others constraint on the weights of an autoencoder to achieve compressions? I want to use the encoder part to replicate a better basis with the STL-10 dataset. Autoencoders | Kaggle The features learned should be more optimized than the compression algorithms like JPEG.
I suggest to take an overview at:
Then you can take a look at the CLIC 21 tasks/leaderboard:
See also our repository at:
Thanks, will look into it. I am curious if I can replace L1 regularization with autoencoder?
The first link two links are specifically related to the image (or video) compression tasks.
The TF compression repository instead it is related to multiple data compression use cases.
I am curious if I can replace L1 regularization with autoencoder?
Generally you can use also
KL-divergence as sparsity penalty.