expose(f, names)moves parameter blocks of a factor into channels, so a graph variable can hold a weight: priors on weights, weight tying and inference over weights become ordinary wiring and message passing. Generic when the blocks are points; exact beliefs on blocks need per-factor methods.
Sources: code:
exposure.jl; the linear-Gaussian methods in the top-levelsrc/exposure.jl; tests in the top-leveltest/exposure.jlTheory (CT-ML wiki): Bayesian Lens · Statistical Game
Implements §1 of L-SPEC-003 Exposed Parameters and Energy Channels. Theory: Unusual Wirings, Hierarchical Models §1.
The wrapper
ExposedFactor holds the factor, the exposed parameter names, their channel names (a block
may be renamed, :A => :w) and the block sizes, read once from initialparameters at
construction. Its channels are f’s plus one per block, of length length(block); the
channel carries vec(block), column-major. Its initialparameters are f’s without the
exposed blocks, so graph-level parameter trees, freezing and optimisers see only what remains.
Energy spaces and scalarisations are forwarded.
A message or free energy goes one of two ways:
| exposed inputs | target | what runs |
|---|---|---|
all DiracBelief | one of f’s channels | f’s own factor_message / local_free_energy, with the blocks substituted into ps (§1.1, every factor) |
| any belief, missing, or the target is a block | any | parameter_message / parameter_free_energy of f, which throw by default |
Polarities and messages handed to f are renamed to its view: own channels as they are,
exposed channels under their parameter names. supports_polarity accepts f’s polarities
with every block Observed(), and whatever supports_parameter_polarity(f, ·) accepts;
supported_polarities adds, for each block, the polarity with that block the target, when
supported.
uses_cavity is true: the substitution reads the incoming messages on the exposed edges.
supported_polarities lists every accepted assignment up to five channels in all; above that
it throws, as L-SPEC-001 permits, because each assignment is a separate Polarity type and
listing of them compiles that many (21.9 s for six channels, measured by Codex,
2026-10-30). hasnopolarity and isunidirectional are exact, by the lazy
LenticulumCore.polarities search stopping at the first (resp. second) accepted assignment;
a structural shortcut was wrong for custom hooks (Codex, 2026-10-30). A search that finds
none or exactly one visits all candidates, so these predicates are slow for factors with
many channels.
A factor with a channel named like an exposed block can be exposed, but only on the clamped
path: parameter_message sees channels and exposed parameters in one namespace, so the
belief path refuses it. A factor whose belief rules read further parameter names checks
those itself (GaussianFactor refuses channels named :A or :b there; :noise is fine).
What it does not do
- Approximate messages. A factor without the three methods supports only points on its blocks; no Laplace or sampling fallback exists. That keeps every result exact or refused.
- Game-level forwarding.
assemble,energy,free_energyandentropyare not defined for the wrapper; it is for graphs. - Nested blocks. Only top-level array entries of
pscan be exposed, so a Lux network’s weights (nestedNamedTuples) cannot yet. - Gradients through exposure. The graph gradient drivers (LP-002, LP-011) are untested on exposed factors.
Difficulties
- A clamped weight whose edge is
Bidirectional()still receives a message back in the tree’s outward pass, and with an uncertain input that message is bilinear and throws. Wire a clamped block like data,Absorbing(). - Leaf predictions add a factor whose output has no incoming message; the factor belief still integrates to one, so the tree’s Bethe free energy stays (tested).