Neural Codebook Channels: A New Diagnostic for VAE Latent Spaces
A new arXiv paper introduces the neural codebook channel, a diagnostic tool for variational autoencoders (VAEs) that measures whether the decoder correctly interprets latent codes under the encoder's codebook. Standard VAE diagnostics like ELBO, active units, mutual information, and code histograms only confirm code usage, not code alignment. The proposed method, K_{e→d}(j|i), quantifies encoder-decoder mismatch with a Bernoulli-KL certificate bounded by the variational gap. This work addresses a fundamental gap in VAE evaluation, with implications for clustering, conditional generation, and mechanistic interpretability.
Key facts
- Paper title: 'Lost and Found in Translation: Variational Diagnostics for Neural Codebook Channels'
- Published on arXiv with ID 2605.18846
- Introduces neural codebook channel K_{e→d}(j|i) for encoder-decoder alignment
- Standard VAE diagnostics (ELBO, active units, mutual information, code histograms) only check code usage, not code alignment
- Proposes a Bernoulli-KL certificate d_bin(1-A || η_p) ≤ Δ controlled by variational gap
- Addresses mismatch when transmitter and receiver use incompatible operational codebooks
- Relevant for clustering, conditional generation, and mechanistic interpretability
- Architecture-free certificate bounds off-diagonal mass of the codebook channel
Entities
Institutions
- arXiv