AI 용어집
인공지능 완전 사전
Factor Disentanglement
Property where each dimension of the latent space captures a single interpretable and independent feature of the data, enabling granular control over generation.
Encoder Capacity
Amount of information that the encoder can transmit to the latent space, indirectly controlled by the beta parameter to prevent overfitting.
Variational Posterior Distribution
Parametric approximation of the true posterior distribution, typically modeled as a diagonal Gaussian whose parameters are learned by the encoder.
Latent Factors of Variation
Independent dimensions of the latent space corresponding to the fundamental underlying attributes that generated the observed data.
Isotropic Prior
Standard Gaussian prior distribution (zero mean, unit variance) used in VAEs to regularize the latent space and facilitate sampling.
Beta Annealing
Training strategy that progressively varies the beta parameter from an initial low value to its target value to improve convergence and disentanglement.
Disentanglement Coefficient
Quantitative metric evaluating the degree to which each latent dimension captures a unique and independent factor of variation in the learned representations.
Amortized Variational Encoding
Process where variational inference is performed by a neural network that learns to directly map data to the parameters of the posterior distribution.
Controlled reconstruction
Beta-VAE's ability to faithfully reconstruct data while maintaining a disentangled latent structure through the trade-off regulated by the beta parameter.
Latent bias-variance trade-off
Fundamental trade-off in Beta-VAE between the capacity of the latent space (variance) and the regularization constraint (bias) towards a simple distribution.
Conditional latent generation
Process of generating new data by selectively manipulating dimensions of the disentangled latent space to control desired attributes.