overview

Section L of the Feature Backlog: M-steps, gradient semantics, train!, hyperparameters, and losses for energy factors. 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
L1The warning of LP-022 for data that arrive other than by DataFactor (pins, point messages)proposed—lowLP-022 Constraint factors are normalised in their residual, not their channels
L2MAP M-steps for noise priors (Wishart, log-normal); M-steps for further familiesproposed—lowlib/Lenticulum.jl/src/m_steps.md §4GaussianFactor and GaussianPrior have ML M-steps since 2026-11-05; with a prior their noise is not updated
L3Stop-gradient semantics per wire, factor and parameter subsetplanned—mediumL-SPEC-002 Stop Gradient
L4Exposed parameters: generic approximate parameter_message, energy taps, risk scalarisationspartial—lowL-SPEC-003 Exposed Parameters and Energy Channels §2–§3
L5Explicit gradient semantics in results (envelope, finite replay, implicit fixed point)proposed—lowChatGPT review §4.2a caller should see which objective was differentiated
L6Calling validate_parameters automatically before inferenceproposed—lowLP-023it exists (2026-11-04) but is opt-in
L7train!: interleave inference sweeps and OptimiserFactor updates, with a check of condition (O) (a rule’s fixed points are the stationary points) for user rulesproposed—mediumThe Regulariser and the Loss §3; lib/Mycelium.jl/src/factors.md §5the optimiser is a factor, but no loop drives it; GradientDescent, Momentum and Nesterov satisfy (O) by their definitions
L8Hyperparameters by the evidence (type-II ML): an M-step for the of normalised priors inside fit_emproposed—mediumThe Regulariser and the Loss §2MacKay’s evidence has an interior optimum when the prior is normalised
L9Hypergradients through fixed points: by the implicit function theorem, through fit_em and the message fixed pointproposed—lowThe Regulariser and the Loss §5the adjoint machinery exists for implicit learners and DEQs; it would make held-out tuning of priors and loss weights a gradient problem
L10Learned scalarisations and loss weights with condition (L) enforced: a warning, as for LP-022, when a learned weight or penalty is fitted on its own objectiveproposed—lowThe Regulariser and the Loss §4derived there: a loss weight or an unnormalised penalty fitted on the training objective goes to zero
L11Amortised inference graph-wide, with the amortisation gap reportedproposed—lowThe Regulariser and the Loss §2; The Inferencer and the Optimizer §7AmortisedInversion exists per factor only
L12Learned optimisers (an update lens with parameters, meta-trained on an outer criterion)proposedL7, L9lowThe Regulariser and the Loss §3research; needs train! and the outer problem first
L13Energy discrepancy as the loss of a general energy factor (toric with normalised = false, DEQ energies, user energies): energies at perturbed data, no MCMC and no score (arXiv:2307.06431)proposed—mediumTraining Energy-Based Models §4, §6.1these factors are only warned about today (LP-022); this would train them correctly. Test: recovers a Gaussian’s energy up to a constant
L14The absolute level of a learned energy: NCE against a known noise, or telescoping density ratios, to fix an EnergyNetwork’s constant per level and modelproposed—mediumTraining Energy-Based Models §6.2needed to compare models and structures by energy, and for a log-evidence beyond the Gaussian case. Test: a Gaussian mixture’s logpartition
L15A time-independent energy for the implicit learner (Energy Matching, arXiv:2504.10612; Equilibrium Matching, arXiv:2510.02300): one field whose roots are the relation and whose descent samplesproposedF2lowTraining Energy-Based Models §4, §6.3replaces the sum over noise levels; compare on the calibration benchmark
L16Sliced score matching for energy factors on clean data (no noise level)proposed—lowTraining Energy-Based Models §4score matching where no diffusion model is wanted
L17Diffusion recovery likelihood (short conditional chains per noise level)proposed—lowTraining Energy-Based Models §4a fallback if denoising score matching proves insufficient
L18For LenticulumLM: a sequence-level residual energy per diffusion step (EDLM, arXiv:2410.21357) and concrete score matching for a discrete implicit learner (arXiv:2211.00802)proposed—lowTraining Energy-Based Models §4the parallel-decoding correction and the discrete learner’s loss