New theory explains spectral instabilities in VAE representation learning

Researchers have developed a new theoretical framework to understand dimension-wise posterior collapse in variational autoencoders (VAEs). By treating the negative evidence lower bound as an effective free energy, they defined a Gaussian theory where the Hessian matrix represents latent fluctuation masses. This approach identifies collapsed directions as an invariant fluctuation sector and derives their mass spectrum using a conditional residual operator. The findings suggest that a local reactivation direction can reduce free energy when decoder variance drops below a specific spectral threshold, marking a point of marginality. AI

RANK_REASON The cluster contains a research paper detailing theoretical advancements in representation learning for variational autoencoders. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• feature learning
• Gaussian function
• Gotit.pub
• Hugging Face
• IArxiv
• principal component analysis
• ScienceCast
• variational autoencoder

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →