denoise autoencoder

• wavenet denoiser is one: https://arxiv.org/abs/1706.07162
• on train time:
  • 1. you feed in your data (make sure it’s not noisy)
  • 2. the noise autoencoder adds noise to your data
  • 3. it tries to predict the original data you fed in (1)
• So during inference time, if your data is noisy, it will remove the noise.
• you CANNOT train a denoise autoencoder if you don’t know the (non-noisy) data
  • cause it doesn’t have a non-noisy signal to act as the ground truth
• Think of denoise autoencoders like normal autoencoders, except you feed in noisy data and it has to reconstruct the non-noisy data