AI-ordlista
Den kompletta ordlistan över AI
Recurrent VAE
Variational Autoencoder architecture integrating recurrent networks to capture temporal dependencies in sequential data, enabling generation and reconstruction of time series.
LSTM-VAE
Variant of Recurrent VAE using LSTM (Long Short-Term Memory) layers in both encoder and decoder to effectively model long-term dependencies in temporal sequences.
GRU-VAE
Recurrent VAE architecture employing GRU (Gated Recurrent Units) to reduce computational complexity while maintaining the ability to model temporal dependencies.
Temporal Latent Space
Temporally structured latent space where each dimension represents evolving characteristics of the sequence, enabling coherent interpolation between temporal states.
Sequential Prior Distribution
Prior distribution in recurrent VAEs that models the probable evolution of latent variables over time, often implemented as a Markovian or Gaussian process.
Temporal Reconstruction Loss
Loss function adapted for sequences that penalizes not only pointwise reconstruction errors but also temporal inconsistencies in generated sequences.
Dynamic Latent Variables
Latent variables that evolve dynamically over time in Recurrent VAEs, capturing the gradual changes in underlying characteristics of sequential data.
Hierarchical Recurrent VAE
Multi-level architecture combining multiple hierarchical recurrent layers to model different temporal scales, from local patterns to global trends.
Attention Mechanism in VAE
Attention mechanism integrated into recurrent VAEs to selectively weight relevant parts of the sequence during encoding and decoding.
Variational Recurrent Neural Network
Probabilistic formulation of recurrent networks where hidden states are treated as random variables with learned distributions, rather than deterministic.
Temporal KL Divergence
Regularization term in Recurrent VAEs that measures the divergence between the temporal posterior distribution and the prior distribution at each time step.
Sequence-to-Sequence VAE
Bidirectional architecture where the encoder reads a complete sequence to produce a latent context, and the decoder generates a new sequence based on this compressed representation.
Temporal Posterior Collapse
Specific problem in Recurrent VAEs where the model ignores temporal latent variables and directly models the data instead, reducing the effectiveness of regularization.
Conditional Temporal Generation
Ability of Recurrent VAEs to generate sequences conditioned on specific inputs or initial states, enabling precise control over generated temporal characteristics.