overview

Section I of the Feature Backlog: compilation, execution plans, schedules, batching, and inference as dynamics on the free energy. Open rows only; a row that is done moves to Backlog Done 2026-11 with its id. Status, priority, ids and the column needs are explained in Feature Backlog, which also holds the dependency graph.

Sources: original to this repository; the sources of each row are in its source column.

idfeaturestatusneedsprioritysourcewhy
I1Compilation per structure in the free-energy and EM pathspartial: message loop (2026-11-04), free energy, cluster code and EM outside AD (2026-11-05)—mediumpassing.md “Compilation per structure”a new structure still compiles 1.14 s for F, em_step and a fit (0.31 s warm); the AD paths and ForwardDiff’s chunks remain
I2Typed execution plans: factors grouped by type and rule, kernels compiled once per signatureproposed—mediumboth reviews (LP-023)the graph stores factors abstractly; plates are the first step
I3A spanning-tree schedule for loopy graphs (spanning_tree_schedule(g): the messages off a spanning tree, then the tree’s inward and outward sweeps, in place; the updates of tree-based reparameterisation)partial: done in Mycelium (2026-11-09; schedules.md); the choice of tree and alternating trees remain—highComposing Optimisers §4.2 (2026-11-09)measured on the divider (bench/bp_schedules.jl): 0.17–0.18 of flooding’s message evaluations to an error of 1e-10, in all four settings, with no knowledge of the loop; with infer! 82 sweeps against 463 (test/runtests.jl). On Gaussian grids up to 20 × 20 (bench/bp_grid.jl, 2026-11-09): 0.17–0.29 of flooding’s sweeps; its Jacobian has at most two nonzero eigenvalues per independent loop (Composing Optimisers §4.4). Untested: graphs where BP oscillates, non-Gaussian graphs
I4A dependency-aware priority queue for ResidualScheduleproposed—mediumChatGPT review (LP-023)each sweep probes every task, evaluations: 2.1 times flooding’s on the divider. A queue prototype in bench/bp_schedules.jl needs 0.24–0.31 of flooding’s, behind the spanning-tree order (Composing Optimisers §4.2)
I7Mixture-reduction and EP nodes batched in platesplanned—lowL-SPEC-008 Expectation Propagation and Mixture Reduction openedge by edge today
I8Clusters with loops; several mean-field variables in one clusterplanned—lowL-SPEC-007 Hybrid Schedules open
I9Nested plates; stateful tied factorsplanned—lowL-SPEC-006 Plates open
I10Incremental inference (reuse factorisations and beliefs after a local change, iSAM2-like)proposed—lowChatGPT review; LenticulumLM’s Incremental Next-Token Prediction (an idea, 2026-11-05)online learning, SLAM-like use, and next-token prediction as appending factors; transfer gives warm starts today, not reuse of factorisations
I11A lazy product-of-potentials belief with explicit approximation policiesproposed—lowChatGPT reviewlearned potentials without silent Gaussian projection
I12Underdamped Langevin dynamics on the Bethe free energy: beliefs (with a mass, e.g. Fisher) get a momentum , with , . At the friction brakes at every step (), a Lyapunov function that a BP sweep lacks; with annealed to 0 it could also hop between fixed points. An idea, untested. To settle: must it descend directly (a BP sweep is not a gradient of , so momentum on messages may amplify oscillation); how to stay on the constraints (normalised, consistent beliefs: mirror or natural coordinates); Gaussian Bethe energies can be unbounded below on loops (from memory, to check). Test: iterations and answers on the loopy graphs that oscillate today, against damped BP and the exact marginals. Interface (checked 2026-11-06): a schedule defined outside the package plugs in through AbstractSchedule and a sweep! method, and infer! runs it (scratch test on the divider graph: at it reproduces flooding exactly; the sweep count it gave is not recorded, bench/bp_spectrum.jl measures 326 sweeps to an error of 1e-10). Missing, all additive: the gradient of in the beliefs (free_energy_gradient is for parameters), coordinates and a retraction for beliefs, and a state holding primal beliefs (off a fixed point the store’s messages do not give consistent beliefs). Naive heavy ball on BP messages failed there for every tried (negative precisions; with momentum on only, NaN)proposedI15lowDaniel, 2026-11-06loopy BP has no convergence guarantee and damping only sometimes cures oscillation (Loopy Message Passing); Momentum exists for parameters but nothing for beliefs
I13Accelerated mirror descent on the Bethe free energy: descend in the beliefs with the entropy (log-partition) as the mirror map, so every iterate is a valid belief by construction (for exponential families, mirror descent in the mean parameters is a natural-gradient step in the natural ones: Raskutti & Mukherjee, arXiv:1310.7780), plus momentum (accelerated mirror descent: Krichene, Bayen & Bartlett, NeurIPS 2015, metadata to check; stochastic version with averaging, the analogue of the Langevin row: Krichene & Bartlett, arXiv:1707.06219)proposedI15, T1lowDaniel, 2026-11-07the failure of naive momentum in the row above was an invalid belief (a negative precision); a mirror map rules that out. Needs the same three interface additions. A complicated optimiser: test it on the benchmark below (Tooling), not on one graph
I14A spectral diagnostic for message passing (partial: Anderson done), and Anderson acceleration as a schedule: report the leading eigenvalues of the linearised sweep at a fixed point (exact for Gaussian means once the precisions converge; AD or finite differences of sweep! otherwise), and an AbstractSchedule that wraps another with Anderson mixing of the last sweepsproposed—mediumDaniel, 2026-11-07 (optimiser composition); Composing Optimisers §4–5the spectrum says in advance whether momentum (real eigenvalues near 1), damping (near −1) or Anderson (complex) should help; heavy ball destabilises complex eigenvalues of modulus 0.95 for every β ≥ 0.3 (closed form). Anderson keeps BP’s fixed points and needs none of the belief-geometry additions above. Measured once (2026-11-08): bench/bp_spectrum.jl computes the spectrum on the divider (all slow eigenvalues complex, on a ring; heavy ball diverges for β ≥ 0.07; a schedule round the loop converges at ρ⁶ per sweep), Composing Optimisers §4.1. Anderson measured (2026-11-09), bench/bp_anderson.jl, §4.3: around spanning_tree_schedule with –3 it needs 0.09–0.19 of flooding’s evaluations, and 414 instead of 297 108 where the loop gain is near 1; around flooding it is worse in the ordinary cases or fails. AndersonSchedule(inner; m) done in Mycelium (2026-11-09, anderson.md): around the spanning-tree order, by infer!, 24 sweeps instead of 20 881 with equal resistors. The diagnostic as a library function remains
I15The belief-space interface that I12 and I13 name as missing: the gradient of in the beliefs (free_energy_gradient is for parameters), coordinates and a retraction for beliefs, and a state holding primal beliefs (off a fixed point the store’s messages do not give consistent beliefs)proposed—lowI12 “Interface (checked 2026-11-06)“split out of I12 and I13 (2026-11-11), which both need it; all additive
I16A Zygote backward pass through the free energy that is linear in the number of parameter keys: a cotangent for ps gathered once per key instead of one per factor readproposedI5mediumI5 design (2026-11-11, Daniel: “put it in the backlog”)expected, not measured: Zygote’s pullback for a read from a NamedTuple, by name or by position, builds a cotangent with an entry for every key, so a gradient through scalar_free_energy stays O(factors × keys) after I5 removes the quadratic forward cost. To measure first: time Zygote.gradient of the free energy on bench/state_scaling.jl’s chain per doubling
I17The coupled gradient drivers and the graph adjoint read by position too (gradient_couplings.jl: _setsubtree per task, ps[name] and _key_node per key; graph_adjoint.jl)proposedI5mediumI5 (2026-11-12, lib/Mycelium.jl/src/passing.md, “The trees by position”)I5 made the envelope gradient and parameter_step linear, not these; expected quadratic for the same reason, not measured