Skip to content

DendriNet 0.3.0

Choose a tag to compare

@houman1359 houman1359 released this 16 Oct 19:09
· 135 commits to develop since this release
8e884ea

Upgrading from 0.2.0

  • Configuration files now carry config_version: 4. Standard dendritic E/I cores are validated as biological neurons unless they declare biological_neuron: false, which stays available for exploratory models.
  • Run dendritic-upgrade-config on configuration files written for earlier versions: it writes the historical biological_neuron: false, so they keep describing the same models and runs. A file without config_version is read as format 4.
  • Synaptic weight normalization (weight_norm_order) now normalizes rows whose norm lies below machine epsilon instead of dividing them by machine epsilon. Rows at or above machine epsilon are unchanged bit for bit, but checkpoints with such small rows can give different outputs, so compare before moving existing results to 0.3.0.

New

  • Checked-identity nonnegative inputs that reject negative or non-finite rates, declared paired_afferents (distinct, identically tuned external E/I sources) and opt-in positive input projections (input_projection_weight_transform: softplus) for population and unified E/I networks.
  • Opt-in shared_projection_shunting core: normalized sparse projections shared across native DendriNet banks, with signed block pooling and an affine output.
  • Biological configuration preparation and a source-pinned recipe validator (dendritic_modeling.scripts.tools.biological_recipes) for migrating experiment recipes.
  • dendritic-collect-reactivation --require-clean-analysis-source rejects runs whose analysis source is not one clean commit.

Fixes and performance

  • Recurrent population and unified E/I networks check nonnegative inputs once per sequence instead of at every step; outputs, recurrent states and gradients are unchanged.
  • Weight normalization below machine epsilon is exact, also when weights are subnormal or the L2 norm underflows; exp and softplus rows keep finite gradients, and the normalization traces as a single graph under torch.compile.
  • Saved 0.2.0 core records stay recognized; flat population and unified E/I payloads accept biological_neuron: false; section-wrapped shared_projection_shunting mappings build; dictionary AlexNet recipes are no longer modified; the reactivation collection configuration rejects unknown keys and mistyped values.
  • Biological operator cards cover the neuronal core only (operator-card schema v3).