Unverified50% confidenceFactExact time
VQ-VAE uses a straight-through estimator to pass gradients back from the decoder to the encoder, and the training objective typically includes three terms: reconstruction loss, codebook alignment loss, and commitment loss.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
Unverified应对损失尖峰的工程策略包括梯度裁剪（将梯度范数限制在1.0以内）、从尖峰前检查点回滚重训以及使用QK Normalization71% similarUnverifiedCodebook collapse is a common failure mode of VQ-VAE, meaning only a few codebook entries are used while the rest become dead codes; common mitigations include exponential moving average updates, codebook resets and diversity regularization.69% similarUnverifiedDuring training, MRL simultaneously optimizes loss functions at multiple dimensional levels, with the total loss being a weighted sum of losses at each granularity.69% similarUnverifiedThe training objective of Diffusion-LM includes three loss terms: the diffusion denoising loss, an embedding matching term, and a rounding loss.69% similarVerifiedAfter geometric transformation, the transformation parameters must be recorded and the predicted boxes must be inverse-transformed to align them; ignoring this step will cause semantic misalignment in the consistency loss and harm training effectiveness68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088745API
curl https://kongchang.com/api/v1/knowledge/claims/1088745MCP
get_claim(id=1088745)