PID analysis pipeline: Gaussian MMI-PID across CTRNN/Elman ensembles, GPU speedup, and cleanup by jgendra · Pull Request #43 · jgendra/NeuroAI-Project13

jgendra · 2026-07-04T21:33:45Z

TL;DR: Adds a full, verified Partial Information Decomposition (PID) analysis pipeline (notebook 06) comparing perceptual vs. context-dependent decision-making RNNs, plus a hidden-size sweep study and repo cleanup (removed an unrelated lecture example, reorganized results/figures directories).

Description:
This branch builds out the PID analysis workflow end-to-end: computing Gaussian MMI-PID over pre-trained CTRNN/Elman ensembles to test whether synergy between stimulus and context is higher for context-dependent decision-making than for purely perceptual decision-making. It includes a GPU-batched PID implementation for tractable runtime, time-resolved and decision-time PID figures across all seeds, accuracy/loss comparison plots, and a separate hidden-size sweep exploring how PID components scale with RNN capacity. Supporting data (trained model weights, activations, stimulus coherences, PID outputs) and output directories were added/reorganized alongside the code. An unrelated old lecture exercise (Jan_example_rnn/) was removed from the repo.

Changes:

notebooks/06_Full_PID_analysis_pipeline.ipynb: new full pipeline — loads the 10-seed × {perceptual, context} CTRNN ensemble, computes Gaussian MMI-PID (total MI, redundancy, unique×2, synergy) against signed stimulus coherence, produces accuracy/loss bar plots with Mann-Whitney U control and time-resolved PID plots for all seeds; iterated across several commits (markdown/comments pass, figure generation, final verification pass).
src/analysis/gaussian_pid.py: new GPU-batched PID implementation (gpu_time_pid) that computes the joint covariance once per timestep and batches sub-block slogdets, ~40x faster than the reference per-seed PID computation, validated to match the reference implementation to 3e-10 bits.
figures/: new accuracy_loss/ (test accuracy/loss bar plot) and all_time_pid/ (time-resolved PID plots per CTRNN seed) directories; added .gitkeep placeholders for learning_curves/{context,perceptual}.
results/: new output directories for accuracies/losses, model activations, model weights, PID outputs (including all_time_PID.npy + metadata for both tasks), and stimulus coherences (test coherence .npz files for context and perceptual tasks, 10 seeds each).
size_comparison/: new hidden-size sweep study for the Context Decision-Making (CDM) and Perceptual Decision-Making (PDM) tasks — sweep script (ctrnn_pid_sweep_cdm.py), PID-vs-hidden-size plots, per-hidden-size PID timeseries plots, and a comparison notebook (plot_pid.ipynb), plus a data-loading verification notebook/script.
notebooks/04_minimal_product.ipynb, notebooks/05_Full_train_pipeline.ipynb: corrected data path references.
Removed notebooks/Jan_example_rnn/ (unrelated Scientific Computing lecture exercise and its saved RNN weights).
.gitignore: updated to ignore Claude-related files and coherence results directories.

…nd 05

…y for elman vs ctrnn pid study, git ignore claude files and coh results

Copilot

Pull request overview

Adds an end-to-end PID analysis workflow to the project, including a GPU-batched Gaussian MMI-PID implementation and supporting scripts/artifacts for comparing perceptual vs. context-dependent RNNs and running hidden-size sweeps.

Changes:

Added a GPU-accelerated, time-resolved Gaussian MMI-PID path for RNN activations.
Added size-comparison / sweep scripts and reorganized result/figure output directories via placeholder folders and metadata.
Cleaned up notebooks/paths and removed an unrelated lecture example; added pre-trained weight artifacts for comparisons.

Reviewed changes

Copilot reviewed 13 out of 135 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
src/training/.gitkeep	Keeps `src/training/` tracked in git.
src/models/.gitkeep	Keeps `src/models/` tracked in git.
src/analysis/gaussian_pid.py	Adds Torch-based GPU-batched time-resolved Gaussian MMI-PID implementation and bipartition helper.
size_comparison/data_loading_ver/ctrnn_pid_both_tasks.py	New CTRNN training + PID analysis script for both tasks with plotting/saving.
size_comparison/CDM/results/ctrnn_pid_sweep_cdm.py	New CDM hidden-size sweep script with PID analysis and plotting.
results/stimulus_coherences/perceptual/.gitkeep	Tracks perceptual coherences output directory.
results/stimulus_coherences/context/.gitkeep	Tracks context coherences output directory.
results/pid_outputs/perceptual/all_time_PID_meta.json	Adds metadata describing saved perceptual all-time PID array.
results/pid_outputs/perceptual/.gitkeep	Tracks perceptual PID outputs directory.
results/pid_outputs/context/all_time_PID_meta.json	Adds metadata describing saved context all-time PID array.
results/pid_outputs/context/.gitkeep	Tracks context PID outputs directory.
results/pid_outputs/.gitkeep	Tracks PID outputs root directory.
results/model_weights/perceptual/.gitkeep	Tracks perceptual model weights output directory.
results/model_weights/context/.gitkeep	Tracks context model weights output directory.
results/model_weights/.gitkeep	Tracks model weights root directory.
results/model_activations/perceptual/.gitkeep	Tracks perceptual activations output directory.
results/model_activations/context/.gitkeep	Tracks context activations output directory.
results/model_activations/.gitkeep	Tracks activations root directory.
results/accuracies_n_losses/perceptual/.gitkeep	Tracks perceptual accuracy/loss output directory.
results/accuracies_n_losses/context/.gitkeep	Tracks context accuracy/loss output directory.
notebooks/Jan_example_rnn/true_rnn/utils.py	Removes old lecture/example utility code.
notebooks/Jan_example_rnn/true_rnn/trueRNN.py	Removes old lecture/example model code.
notebooks/Jan_example_rnn/true_rnn/main.py	Removes old lecture/example training script.
notebooks/Jan_example_rnn/dyn_rnn/utils.py	Removes old lecture/example utility code.
notebooks/Jan_example_rnn/dyn_rnn/main.py	Removes old lecture/example training script.
notebooks/Jan_example_rnn/dyn_rnn/dynRNN.py	Removes old lecture/example model code.
notebooks/05_Full_train_pipeline.ipynb	Updates base data paths and clears execution counts / kernel metadata.
notebooks/04_minimal_product.ipynb	Updates base data paths, clears execution counts, and adjusts weights output path.
figures/learning_curves/perceptual/.gitkeep	Tracks perceptual learning-curve figure directory.
figures/learning_curves/context/.gitkeep	Tracks context learning-curve figure directory.
figures/learning_curves/.gitkeep	Tracks learning-curves root directory.
elman_vs_ctrnn_comparison/model_weights/Elman_PerceptualDecisionMaking-v0_h80.pt	Adds pre-trained Elman weights artifact.
elman_vs_ctrnn_comparison/model_weights/Elman_PerceptualDecisionMaking-v0_h20.pt	Adds pre-trained Elman weights artifact.
elman_vs_ctrnn_comparison/model_weights/Elman_ContextDecisionMaking-v0_h20.pt	Adds pre-trained Elman weights artifact.
elman_vs_ctrnn_comparison/model_weights/CTRNN_PerceptualDecisionMaking-v0_h20.pt	Adds pre-trained CTRNN weights artifact.
elman_vs_ctrnn_comparison/model_weights/CTRNN_ContextDecisionMaking-v0_h20.pt	Adds pre-trained CTRNN weights artifact.
.gitignore	Adjusts ignore/unignore rules for results coherences and ignores `.claude/`.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+    Fast drop-in for `gaussian_pid_rnn(..., timestep=None, bipartitions="random")`:
+    it returns the same bipartition-averaged PID atoms at every timestep, but is
+    ~40x faster. The trick is that at each timestep the joint covariance over *all*


+
+
+
+def gaussian_pid_rnn_gpu(H, Y, n_bip=200, seed=0, reg=1e-5, device=None):


+RESULTS_DIR = Path(r"E:\TUM\3sem\NeuroAI\project\github_neuroai\results")
+
+WEIGHTS_DIR  = RESULTS_DIR / 'weights'
+WEIGHTS_DIR.mkdir(parents=True, exist_ok=True)
+


+        test_acc = compute_decision_accuracy(model, loaders[task_key]['test'], device)
+        history['test_acc'] = test_acc
+        print(f"  Test decision accuracy: {test_acc:.3f}")


+for h in args.hidden_sizes:
+    print(f"\n{'#'*60}")
+    print(f"  HIDDEN SIZE = {h}")
+    print(f"{'#'*60}")
+
+    main(
+        hidden_size    = h,
+        pdm_dir        = args.pdm_dir,
+        cdm_dir        = args.cdm_dir,
+        num_epochs     = args.num_epochs,
+        patience       = args.patience,
+        lr             = args.lr,
+        batch_size     = args.batch_size,
+        n_test_trials  = args.n_test_trials,
+        n_bipartitions = args.n_bipartitions,
+    )


+    'fixation': 200,    # 10 steps @ dt=20
+    'stimulus': 1000,    # 50 steps (rounded to 38 at dt=20)
+    'delay':    0,      # 0 steps  — MUST fix stochastic delay
+    'decision': 100,    # 5  steps


jgendra added 8 commits July 4, 2026 15:41

Add Harris RNN size exploration, corrected data path of notebook 04 a…

62afbaf

…nd 05

correct data path notebooks 04 & 05, add coherences to results

487fe4d

perceptual coherences in results

bca343e

PID analysis acc loss bar plot, time resolved pid plots, new director…

feef038

…y for elman vs ctrnn pid study, git ignore claude files and coh results

full pid pipeline markdown and comments

c8127c5

Time resolved PID figures for all RNNs

7ee98bb

removed jan's example rnn from scientific computing lecture

d3b241e

verified working pid pipeline v1

3faccec

Copilot AI review requested due to automatic review settings July 4, 2026 21:33

jgendra self-assigned this Jul 4, 2026

Copilot started reviewing on behalf of jgendra July 4, 2026 21:34 View session

jgendra merged commit e8a44fb into main Jul 4, 2026
1 check passed

Copilot AI reviewed Jul 4, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

PID analysis pipeline: Gaussian MMI-PID across CTRNN/Elman ensembles, GPU speedup, and cleanup#43

PID analysis pipeline: Gaussian MMI-PID across CTRNN/Elman ensembles, GPU speedup, and cleanup#43
jgendra merged 8 commits into
mainfrom
jgendra

jgendra commented Jul 4, 2026

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants




		def gaussian_pid_rnn_gpu(H, Y, n_bip=200, seed=0, reg=1e-5, device=None):

Conversation

jgendra commented Jul 4, 2026

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants