A denoising autoencoder trains with noisy data in order to create a model able to reduce noise in reconstructions from input data

autoencoder_denoising(network, loss = "mean_squared_error",
  noise_type = "zeros", ...)



Layer construct of class "ruta_network"


Loss function to be optimized


Type of data corruption which will be used to train the autoencoder, as a character string. Available types:

  • "zeros" Randomly set components to zero (noise_zeros)

  • "ones" Randomly set components to one (noise_ones)

  • "saltpepper" Randomly set components to zero or one (noise_saltpepper)

  • "gaussian" Randomly offset each component of an input as drawn from Gaussian distributions with the same variance (additive Gaussian noise, noise_gaussian)

  • "cauchy" Randomly offset each component of an input as drawn from Cauchy distributions with the same scale (additive Cauchy noise, noise_cauchy)


Extra parameters to customize the noisy filter:

  • p The probability that each instance in the input data which will be altered by random noise (for "zeros", "ones" and "saltpepper")

  • var or sd The variance or standard deviation of the Gaussian distribution from which additive noise will be drawn (for "gaussian", only one of those parameters is necessary)

  • scale For the Cauchy distribution


A construct of class "ruta_autoencoder"


  • Extracting and composing robust features with denoising autoencoders

See also