AI 용어집
인공지능 완전 사전
M1 Model
First semi-supervised model using a VAE for unlabeled data and a separate classifier for labeled data, optimized independently.
M2 Model
Improved architecture where the label is integrated as a conditional latent variable, enabling controlled data generation and unified classification.
Joint Optimization
Strategy for simultaneous optimization of the encoder, decoder, and classifier using both labeled and unlabeled data.
Latent Variable Supervision
Technique where labels provide direct supervision on the latent space to guide the learning of discriminative representations.
Hybrid Learning Objective
Loss function combining VAE reconstruction, KL regularization, and classification loss, weighted according to data type.
Classifier Head
Classification module attached to the VAE encoder that predicts labels from the latent representation, trained on labeled data.
Semi-supervised ELBO
Variant of the evidence lower bound adapted for partially labeled data incorporating classification terms.
Representation Disentanglement
Property where the latent space naturally separates semantic variation factors from style factors, facilitated by partial supervision.
Teacher-student VAE
Architecture where a teacher VAE supervises a student VAE to improve the stability of semi-supervised learning.
Variational Semi-supervised Learning
Paradigm combining variational inference with partially supervised data for unified probabilistic modeling.
Latent Classifier
Classifier operating directly in the VAE latent space, leveraging learned representations for better generalization.
Auxiliary Task Learning
Multi-task learning where reconstruction serves as an auxiliary task to improve the main classification performance.