How alexvoce/cortico-cerebellar_rnns works
Cortico-cerebellar RNNs trains recurrent models, optionally augmented with a cerebellar-inspired bias, on temporal-memory tasks and analyzes their learning and mechanisms. The principal workflow runs from task selection and sequence generation through model training, curriculum updates, and saved run artifacts. Separate analysis branches cover learning comparisons and mechanistic studies. The README establishes the product capabilities and training entry point; source-index references and excerpts support the implementation links shown. Some analysis relationships are documented at capability level because their run data and wiring were not fully sampled.
Main entry points: train.py Training entry · model_stepping.py Model dynamics · reservoir.py Reservoir curriculum · parity_dynamics.py Parity dynamics
Components 34
Tasks and curricula
- Task specifications
tasks/registry.py - Sequence generators
tasks/task_generators.py - Multitask training
tasks/multitask.py - Task switching
tasks/task_switch.py - Continual training
tasks/continual.py
Training and runs
- Training entry
training/train.py - Shared train loop
training/base.py - Reservoir curriculum
training/reservoir.py - Reservoir stages
training/variants.py - Optimizer utilities
training/utils.py - Run artifacts
training/save.py
Learning analyses
- Single-task comparisons
analysis/single_task_analysis_utils.py - Learning plots
analysis/single_task_plotting_utils.py - Learning metrics
analysis/learning_metrics.py - Model scaling
analysis/scaling.py - Multitask comparisons
analysis/multitask_switching_plotting_utils.py
Mechanism analyses
- CB ablation
analysis/cb_ablation.py - Model reconstruction
analysis/rebuild_model_utils.py - Model dynamics
analysis/model_stepping.py - DMS dynamics
analysis/dms_dynamics.py - DMS propagation
analysis/dms_memory_propagation.py - Jacobian analysis
analysis/jacobian.py - Parity dynamics
analysis/parity_dynamics.py - Parity update rules
analysis/parity_update_rule.py - Predictive analysis
analysis/predictive_analysis.py - CB encoding
analysis/parity_cb_encoding.py - Readout recovery
analysis/readout_recovery.py - State dimensionality
analysis/pca_dimensionality.py - Timescale analysis
analysis/timescales.py - Mechanism figures
analysis/mechanism_plots.py
Other components
- Researcher
- Trained runs
Connections 43
- Researcher → Training entry: launches
README.md - Training entry → Task specifications: selects task
training/train.py - Task specifications → Sequence generators: uses batches
tasks/registry.py - Training entry → CB-RNN: configures model
training/train.py - Training entry → CB-GRU: configures model
training/train.py - CB-GRU → CB-RNN: reuses CB bias
model/cb_gru.py - Training entry → Multitask training: dispatches mode
training/train.py - Training entry → Task switching: dispatches mode
training/train.py - Training entry → Continual training: dispatches mode
training/train.py - Training entry → Reservoir curriculum: dispatches mode
training/train.py - Reservoir curriculum → Reservoir stages: uses stages
training/reservoir.py - Reservoir curriculum → Optimizer utilities: uses optimizers
training/reservoir.py - Shared train loop → Task specifications: reads specs
training/base.py - Shared train loop → Optimizer utilities: uses utilities
training/base.py - Training entry → Run artifacts: saves runs
training/train.py - Reservoir curriculum → Run artifacts: saves runs
training/reservoir.py - Run artifacts → Trained runs: writes artifacts
README.md - Trained runs → Single-task comparisons: supplies runs
README.md - Trained runs → Learning metrics: supplies runs
README.md - Single-task comparisons → Learning plots: supports figures
README.md - Learning plots → Model scaling: provides plotting
analysis/scaling.py - Trained runs → Multitask comparisons: supplies runs
README.md - Trained runs → CB ablation: supplies runs
README.md - CB ablation → Model reconstruction: uses reconstruction
analysis/cb_ablation.py - Model dynamics → Model reconstruction: rebuilds models
analysis/model_stepping.py - Model dynamics → Sequence generators: generates trials
analysis/model_stepping.py - DMS dynamics → Model dynamics: rolls out models
analysis/dms_dynamics.py - DMS propagation → Model dynamics: uses dynamics
analysis/dms_memory_propagation.py - DMS propagation → DMS dynamics: uses metrics
analysis/dms_memory_propagation.py - Jacobian analysis → Model dynamics: uses dynamics
analysis/jacobian.py - Parity dynamics → Model dynamics: uses dynamics
analysis/parity_dynamics.py - Parity dynamics → Predictive analysis: uses prediction
analysis/parity_dynamics.py - Parity update rules → Model dynamics: uses dynamics
analysis/parity_update_rule.py - Parity update rules → Parity dynamics: uses trials
analysis/parity_update_rule.py - Predictive analysis → CB ablation: uses ablation data
analysis/predictive_analysis.py - CB encoding → Parity dynamics: uses dynamics
analysis/parity_cb_encoding.py - CB encoding → Parity update rules: uses update analysis
analysis/parity_cb_encoding.py - CB encoding → Jacobian analysis: uses Jacobians
analysis/parity_cb_encoding.py - Readout recovery → CB ablation: uses ablation
analysis/readout_recovery.py - State dimensionality → Model reconstruction: rebuilds models
analysis/pca_dimensionality.py - Trained runs → State dimensionality: supplies runs
README.md - Trained runs → Timescale analysis: supplies runs
README.md - Trained runs → Mechanism figures: supplies runs
README.md