AI-ordlista
Den kompletta ordlistan över AI
Salt and Pepper Noise
A form of impulse noise that affects images by replacing certain pixels with extreme values (black or white), used to test the robustness of denoising autoencoders.
Data Corruption
The process of intentionally altering input data by adding noise, serving as input signal to the denoising autoencoder for training.
Robust Latent Space
The compressed representation of data in the denoising autoencoder, designed to be insensitive to noise-induced variations and capture the intrinsic characteristics of the data.
Denoising Autoencoder (DAE)
The English name and common acronym for the denoising autoencoder, a fundamental model in unsupervised learning for regularization.
Denoising Regularization
A regularization technique where training to denoise forces the model to learn general features rather than memorizing training data.
Overcompleteness
A characteristic where the encoding layer of a denoising autoencoder has a dimension higher than that of the input, allowing the model to capture richer representations and better handle noise.
Factor Disentanglement
The ability of a trained denoising autoencoder to separate semantic variation factors of the data from those related to noise in its latent representation.
Dropout Noise
The use of dropout technique as a form of structural noise applied to network activations during training, acting as an effective regularizer.
Sparse Encoding
A constraint applied to the latent space of a denoising autoencoder to activate only a small number of neurons, promoting the learning of more discriminative and noise-robust features.
Stacked Autoencoder
A deep denoising autoencoder architecture composed of multiple layers, enabling the learning of feature hierarchies for more complex and effective denoising.