How allthingssecurity/informationbottleneck works
This is a runnable MNIST research experiment, not a general-purpose information-bottleneck library. A researcher launches the experiment, which loads MNIST and sweeps autoencoder bottleneck sizes; trained models are evaluated and their hidden-layer ReLU regions analyzed for information and compression metrics. The runner then emits plots, tabular results, and a report. The graph preserves the README’s classifier-based prediction measures and the major analysis outputs. Some experiment calls are visible only in partial excerpts or the SOURCE INDEX; edges use the experiment file or README where those relationships are documented.
Main entry points: experiment.py Experiment runner · report.py Study report · models.py Autoencoder and head · mnist_info_region_study Result tables
Components 18
Experiment entry
- Experiment CLI
scripts/run_experiment.py - Experiment runner
src/infobottleneck/experiment.py
Data and modeling
- MNIST dataset
- MNIST loaders
src/infobottleneck/data.py - Autoencoder and head
src/infobottleneck/models.py - Model training
src/infobottleneck/experiment.py - Test evaluation
src/infobottleneck/experiment.py
Information analysis
- Activation capture
src/infobottleneck/analysis.py - Region and MI metrics
src/infobottleneck/analysis.py - Discrete metric helpers
src/infobottleneck/utils.py - Metric correlations
src/infobottleneck/analysis.py
Results and interpretation
- Analysis plots
src/infobottleneck/plots.py - Study report
src/infobottleneck/report.py - Result tables
results/mnist_info_region_study - Model checkpoints
results/mnist_info_region_study/models - Generated figures
results/mnist_info_region_study/plots - Generated report
results/mnist_info_region_study/REPORT.md
Other components
- Researcher
Connections 21
- Researcher → Experiment CLI: launches
README.md - Experiment CLI → Experiment runner: invokes
scripts/run_experiment.py - Experiment runner → MNIST loaders: builds loaders
src/infobottleneck/experiment.py - MNIST loaders → MNIST dataset: loads
src/infobottleneck/data.py - Experiment runner → Model training: runs sweep
README.md - Model training → Autoencoder and head: trains
src/infobottleneck/experiment.py - Experiment runner → Test evaluation: evaluates models
src/infobottleneck/experiment.py - Test evaluation → Autoencoder and head: evaluates
src/infobottleneck/experiment.py - Experiment runner → Model checkpoints: saves models
src/infobottleneck/experiment.py - Experiment runner → Activation capture: collects activations
src/infobottleneck/experiment.py - Activation capture → Autoencoder and head: captures layers
src/infobottleneck/analysis.py - Experiment runner → Region and MI metrics: builds metrics
src/infobottleneck/experiment.py - Region and MI metrics → Discrete metric helpers: summarizes IDs
src/infobottleneck/analysis.py - Experiment runner → Metric correlations: computes correlations
src/infobottleneck/experiment.py - Metric correlations → Result tables: writes metrics
- Experiment runner → Result tables: writes tables
src/infobottleneck/experiment.py - Experiment runner → Analysis plots: generates plots
src/infobottleneck/experiment.py - Analysis plots → Generated figures: saves figures
src/infobottleneck/plots.py - Experiment runner → Study report: generates report
src/infobottleneck/experiment.py - Study report → Result tables: reads tables
src/infobottleneck/report.py - Study report → Generated report: writes report
src/infobottleneck/report.py