Section U of the Feature Backlog: data loading and the small exact factors written by hand today (asked for 2026-11-05). 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.
Daniel: “there should be a whole bunch of utility factors that are still lacking.” Mycelium has
DataFactor, PriorFactor, LossFactor, RelayFactor (an equality between two variables),
OptimiserFactor, TableFactor and StraightThroughFactor. Today the operations below are
written by hand, from LinearConstraintFactors with selection matrices or from graph code.
| id | feature | status | needs | priority | source | why |
|---|---|---|---|---|---|---|
| U1 | Data loading from MLUtils.jl (DataLoader): minibatches fed into a plate’s DataFactors, shuffling, epochs, as an optional extension | proposed | — | medium | Daniel, 2026-11-05 | every example builds one DataFactor per point by hand and keeps all data in the graph; minibatch training needs the plate’s data swapped per batch |
| U2 | A split / concatenate factor: one channel of dimension related exactly to two channels of and (and more pieces, any index sets), with exact messages in every direction | proposed | — | medium | Daniel, 2026-11-05 | today a selection-matrix LinearConstraintFactor with tiny noise, which is neither exact nor cheap. Since 2026-11-11 a HardConstraintFactor with selection matrices imposes it exactly (not tried for this use); a dedicated factor would still be cheaper |
| U3 | A normalising change of variables: an affine factor (per coordinate or per group) with exact messages both ways and the log-Jacobian in its free energy; fixed , from the data (standardising data channels) or learned with data-dependent initialisation (Glow’s ActNorm) | proposed | — | medium | Daniel, 2026-11-05 (batchnorm and group norm in deep-learning libraries) | conditioning: the flat-direction threshold, eigenvalue floors and EM are relative to scale. Without the log-Jacobian a learned scale collapses like LP-022 (squash , raise the density of ). Test: a chain with and without it differs by exactly (mind the deterministic factor’s Bethe bookkeeping, as for the relay in LP-019) |
| U4 | Shared plate statistics: a constructor for a plate whose replicas share mean and scale variables (), the probabilistic reading of batchnorm’s batch statistics | proposed | — | low | Daniel, 2026-11-05 | expressible today (lib/Lenticulum.jl/examples/hierarchical_models.jl); the posterior over replaces running averages at test time. Not batchnorm’s forward pass, which couples every replica through a non-local factor and is no density |
| U5 | Further utility factors: a deterministic map (a function as an exact relation, FunctionFactor), a permutation or reshape of a channel, a sum or average of several channels, a constant (clamp) factor with its own schedule | proposed | — | low | Gemini review §4 (a FunctionFactor), Daniel, 2026-11-05 | the common glue of a model, written by hand each time |