Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Inference Posterior
Probability distribution over latent variables conditioned on observed data, approximated by the VAE encoder because its exact calculation is often intractable.
Latent Space Prior
A priori probability distribution chosen for latent variables, typically a standard isotropic Gaussian N(0, I), which serves as a regularizer and facilitates sampling for generation.
Posterior Collapse
Training problem where the encoder produces latent distributions almost identical to the prior, ignoring input data and making the decoder unable to use latent information for generation.
Conditional VAE
Variant of VAE where generation is conditioned on additional information, such as a class label or attributes, integrated into the encoder and decoder for targeted control.
Denoising VAE
Extension of VAE trained to reconstruct clean data from versions corrupted by noise, improving the robustness of latent representation and the model's generalization capability.
Latent Factor Disentanglement
Objective aiming for each dimension of the latent space to encode a semantically independent and interpretable factor of data variation, facilitating control and interpretability of generation.
Hierarchical VAE
Multi-level VAE architecture where the latent space is structured in layers, capturing abstractions at different scales and improving modeling of complex data distributions.
Amortized Variational Inference
Method where a neural network (the encoder) learns to approximate the calculation of the posterior distribution for any input data, avoiding costly optimization for each sample.