is a type of ANN architecture used to learn (efficient encodings / latent space) of unlabeled data. They do this by taking a high-dimensional signal 𝑥 and compressing it into “latent space” and then decompressing it to a high-dimensional signal 𝑥̅, which is somewhat equivalent to the original signal 𝑥
is a type of manifold learning
is a form of compression
can reconstruct the Principal Component Analysis (PCA)

AE - Introduction

bottleneck hidden layer - forces network to learn a compressed latent representation
reconstruction loss - forces the compressed latent representation to capture/encode as much information about the data as possible

AE - Latent Space Dimensionality → Reconstruction Quality

Beta Variational Autoencoders (𝛽-VAE)	are variational autoencoders but with a modified loss function 𝐿(𝜙, 𝜃, 𝑥) = (𝑟𝑒𝑐𝑜𝑛𝑠𝑡𝑟𝑢𝑐𝑡𝑖𝑜𝑛 𝑙𝑜𝑠𝑠) + 𝛽 * (𝑟𝑒𝑔𝑢𝑙𝑎𝑟𝑖𝑧𝑎𝑡𝑖𝑜𝑛 𝑡𝑒𝑟𝑚) 𝛽>1 - constrains latent bottle neck, encouraging efficient latent encoding → disentanglement Link to original
Convolutional Autoencoders (CAE)	TODO Link to original
Masked Autoencoders (MAE)	a self-supervised representation learner that reconstructs missing parts but is not a generator Link to original
Sparse Autoencoders (SAE)	is a type of neural network used for unsupervised feature learning and dimensionality reduction it is a variant of a standard autoencoder, but with an added sparsity constraint on the hidden layer activations Link to original
Stacked Autoencoders (SAE)	TODO Link to original
Variational Autoencoders (VAE)	is a probabilistic twist on traditional autoencoder - samples the mean and standard deviation to compute latent sample is an ANN architecture introduced by Diederik P. Kingma and Max Welling it belongs to the family of probabilistic graphical models and variational Bayesian methods Link to original