KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Hierarchical Encoder
Part of the deep autoencoder that progressively reduces data dimensionality through multiple layers, capturing increasingly complex abstractions.
Hierarchical Decoder
Part of the deep autoencoder that reconstructs the original data from the compressed latent representation, reversing the encoder process layer by layer.
Deep Latent Space
Low-dimensional compressed representation of data, learned by the central layers of the deep autoencoder, where the most important features are encoded.
Bottleneck
Central layer of the autoencoder with the lowest dimensionality, forcing the network to learn the most concise and informative representation of the data.
Layer-wise Pre-training
Weight initialization technique for a deep autoencoder by sequentially training each encoder-decoder pair as a shallow autoencoder, improving convergence.
Deep Denoising Autoencoder
Deep autoencoder variant trained to reconstruct clean data from noise-corrupted versions, promoting the learning of robust features.
Deep Variational Autoencoder (VAE)
Deep autoencoder where the latent space is constrained to follow a probabilistic distribution (typically Gaussian), enabling new data generation and interpolation.
Sparsity Regularization
Technique adding a penalty to the cost function to encourage neurons in the latent layer to be mostly inactive, promoting more discriminative representations.
Deep Convolutional Autoencoder
Deep autoencoder architecture using convolution and pooling layers to efficiently process structured data like images, capturing spatial patterns.
Factor Disentanglement
Advanced objective aiming for each dimension of the latent space in a deep autoencoder to encode an independent and interpretable semantic variation factor.
Deep Contractive Autoencoder
Deep autoencoder penalized for being insensitive to small variations in input data, promoting the learning of stable and generalizable representations.
Glorot/Xavier Initialization
Method for initializing neuron weights in a deep autoencoder, crucial for avoiding vanishing or exploding gradient problems during training.