Reweighted wake-sleep
Algorithms AIMech:ReweightedWakeSleep DRAFT
Aliases: RWS
A learning procedure for paired generative and inference networks that uses multiple inference samples and normalized importance weights for generative-model updates, with recognition updates from importance-weighted observed-data samples, model-generated sleep samples, or both.
Identity Notes & Concept Scope
This is Bornschein and Bengio’s ICLR 2015 RWS procedure, scoped to its tractable joint/proposal calculations and Algorithm 1. It is distinct from ordinary wake-sleep, the generic importance-sampling identity, BBVI score-function ELBO optimization, and the choice of SBN or NADE component architecture. A single generative wake sample recovers the original wake model update; reproducing original wake-sleep additionally uses the sleep recognition update. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 340 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in worktree and main, returning 24 and 24 lines respectively. All 17 unique matched line contents were inspected in full: NVIL method distinctions, source-reading scopes, prior reviews and the carried wake-sleep discovery lead, not existing candidate records. Exact checkout counts and differences are retained in the ledger. Nine exact AIO CLASS nodes with scoped synonyms/comments and outgoing edges and six full native neighboring records were inspected. Absence is bounded to this snapshot and semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
For an observed input x and fixed current parameters, draw K independent latent configurations h_k from the recognition distribution q_phi(h|x), evaluate p_theta(x,h_k) and q_phi(h_k|x), and compute weights w_k=p_theta(x,h_k)/q_phi(h_k|x). Normalize them by their positive sum. Update theta using the weighted average of grad_theta log p_theta(x,h_k). The unnormalized average of w_k estimates p_theta(x); taking its logarithm gives a distinct, generally biased log-likelihood estimator. For recognition training, the wake option uses the same normalized weights to average grad_phi log q_phi(h_k|x), targeting KL(p_theta(h|x)||q_phi(h|x)) on observed inputs. The sleep option samples a fresh (x-prime,h-prime) from p_theta and ascends log q_phi(h-prime|x-prime). Algorithm 1 permits either recognition option or both. These displayed updates weight log-probability scores evaluated at sampled states; they do not differentiate through the normalized weights as an extra part of that score update. With K=1 the model update becomes the ordinary generative wake update, while the wake recognition score alone has zero expectation under normalized q at fixed parameters, as a curator consequence of the score identity.
⚠️ Curated Limitations & Negative Results
- Importance estimation requires proposal support wherever the target joint is positive, accessible joint/proposal values and suitable integrability. Arbitrary directed or deep-belief models with unavailable joint normalization are not covered merely by their architecture name.
- The likelihood average can be unbiased while its logarithm and normalized-weight gradient remain biased at finite K. A consistent limit under suitable conditions is not an assertion of monotonic finite-K variance or bias improvement, global convergence, or exact posterior inference.
- A proposal equal to the posterior makes likelihood weights constant, but it does not automatically minimize the variance of every model-gradient estimator. Poor proposal coverage can concentrate weights and make the estimate unreliable.
- The recognition wake and sleep options optimize posterior-to-recognition matching under different input distributions. Neither is automatically the recognition gradient of the standard data ELBO used by NVIL, and differentiating an importance-weighted bound with respect to proposal parameters is a different update.
- The main-text prose reverses the arguments of q in one loss expression, while Equations (4)–(5) and Algorithm 1 consistently use q(h|x). Supplement Equation (11) begins with a joint-weighted expression before introducing a conditional normalization; Equation (12) gives the intended posterior-KL target. This record follows those explicit update formulas and does not certify the printed intermediate equalities or author implementation.
- K, component layers and update options are method choices rather than a universal best configuration. No benchmark advantage, author-code execution or runtime scaling claim is included.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:BayesianNetwork
Bayesian Network
|
related_match | The core procedure explicitly learns layered directed probabilistic models using joint evaluations and a recognition proposal. The pinned target is the directed model class, whereas RWS is a parameter-learning procedure with importance weighting. This related mapping does not imply equivalence, subclass membership, or applicability to every model without the required tractable probabilities. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Jörg Bornschein, Yoshua Bengio
|
primary paper
full_text
|
ID: arXiv:1406.2751v4
sha256:ddd18592df0b…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T17:11:00Z
Updated: 2026-10-11T17:11:00Z
Source File:
data/curated/ReweightedWakeSleep.yaml