Denoising autoencoder is using noised added training samples to predict (original) training samples themselves. The goal is to denoise when being applied to the real sample. Here is an example of illustration in this tutorial.
Since the noises are being added to each training sample randomly. I wonder does it make sense to oversample the training set by adding different noise to each training sample? For example,
- duplicate each training sample two times.
- for these two duplicated training samples, randomly add Gaussian noise to each of them respectively.
By doing so, the number of training samples doubled. Then fit the autoencoder.
Does this make sense? Will it improve the reconstruction performance on the real sample?
I have a training set and a test set, where the test set is with noise, and the training set is clean. By doing the above procedure, I surprisingly found the more oversampling, the better the reconstruction performance on the testing set.