CVAE Objective and Loss Terms
CVAE Objective and Loss Terms
This note centralizes the loss-level notes that were already natural atomic units in the original methodology thread.
Core Objective
The base training picture is the conditional ELBO:
The project’s methodology notes mainly study why this objective is under-specified in the entangled-initialized regime, and how to alter it so that carries strategy semantics rather than collapsing.
Across the reconstruction variants, a useful framing is that plain reconstruction measures absolute improvement in predicting , while base-relative normalization turns that into fractional improvement relative to the z-free model.
Atomic Loss Notes
- Reconstruction weighting and normalization Unifies the global, example-level, and token-level reconstruction variants and their interpretations.
- Token-sum versus token-mean NLL aggregation Separates formal sequence likelihood from the supervised-token mean NLL currently used in the implementation, and records why denominators change example and token weighting when lengths vary.
- Distribution-Level Z-Free-Relative Reconstruction Rescaling Uses the entangled baseline loss to restore reconstruction scale globally.
- Example-level inverse-surprisal weighting Explores per-example inverse-surprisal weighting and its example-adaptive effective beta interpretation.
- Token-level surprisal-share weighting Focuses reconstruction on strategy-sensitive teacher-forced positions with high base-model surprisal shares.
- Inter-latent divergence Adds an exclusivity signal so different latents induce different predictive distributions.
- InfoVAE connections Reframes the ELBO through the mutual-information / marginal-matching lens.
- Router marginal KL Describes the batch-level geometric regularizer on aggregate router prior usage.
- Posterior variance Describes the absolute posterior-sharpness diagnostic and auxiliary penalty.
- Sequence-level divergence Records a harder full-sequence divergence direction and why it remains deferred.
How To Read This Cluster
Start with problem setup and model components. Then read reconstruction weighting and normalization for the unifying picture. Then read distribution-level z-free-relative reconstruction rescaling, example-level inverse-surprisal weighting, and token-level surprisal-share weighting. Read inter-latent divergence and InfoVAE connections as the main exclusivity / information-flow add-ons.
