KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Corrupted Input
Input data intentionally degraded by adding noise or masking, used to improve the generalization capacity of autoencoders in real conditions.
Robust Feature Learning
Process of learning stable and invariant features against variations and corruptions of input data.
Masking Noise
Type of noise where certain dimensions of the input are randomly set to zero, simulating missing or incomplete data.
Gaussian Noise
Additive noise following a normal distribution, commonly used to corrupt inputs in denoising autoencoders.
Overcomplete Representation
Latent space with higher dimension than the original input space, allowing to capture richer and redundant features.
Tied Weights
Sharing of weights between encoder and decoder layers, symmetrically transposed to ensure optimal reconstruction.
Corrupted Data Distribution
Statistical distribution of data after applying the corruption process, used to train the model to be robust to perturbations.
Dropout Denoising
Application of dropout as a noise mechanism during training, combining regularization and denoising learning.
Invariant Feature Extraction
Extraction de caractéristiques restant stables malgré les transformations et corruptions appliquées aux données d'entrée.