How pp-00-pp/machine-unlearning-trustworthy-ai works
This repository is a CIFAR-10 machine-unlearning experiment, not a general-purpose application. Researchers create a controlled forget set, train comparison models, apply NegGrad+ to an original model, evaluate forgetting and retained utility, then compare results against retraining and visualize trade-offs. The scripts are standalone; the excerpts show artifact reads and writes in some stages, but not a complete invocation pipeline, so unverified handoffs are omitted or marked dashed.
Main entry points: results Experiment artifacts · unlearn_neggrad_plus.py NegGrad+ unlearning · create_corruption.py Create corruption · evaluate_original.py Original evaluation
Components 14
Data preparation
- Create corruption
src/create_corruption.py
Model experiments
- Baseline training
src/train_baseline.py - Corrupted training
src/train_corrupted.py - Reference retraining
src/train_retrained.py - NegGrad+ unlearning
src/unlearn_neggrad_plus.py
Evaluation and comparison
- Original evaluation
src/evaluate_original.py - Unlearned evaluation
src/evaluate_unlearned.py - Reference evaluation
src/evaluate_retrained.py - Behavior comparison
src/compare_models.py
Results and reporting
- Experiment artifacts
results - Result summary
src/summarize_results.py - Trade-off plots
src/plot_results.py
Other components
- Researcher
- CIFAR-10
Connections 24
- Researcher → Create corruption: runs experiment
README.md - Researcher → Baseline training: runs experiment
README.md - Researcher → Corrupted training: runs experiment
README.md - Researcher → Reference retraining: runs experiment
README.md - Researcher → NegGrad+ unlearning: runs experiment
README.md - CIFAR-10 → Create corruption: provides labels
src/create_corruption.py - Create corruption → Experiment artifacts: writes forget data
src/create_corruption.py - CIFAR-10 → Corrupted training: provides dataset
src/train_corrupted.py - CIFAR-10 → Reference retraining: provides dataset
src/train_retrained.py - Experiment artifacts → NegGrad+ unlearning: supplies forget data
src/unlearn_neggrad_plus.py - Experiment artifacts → NegGrad+ unlearning: supplies original model
src/unlearn_neggrad_plus.py - NegGrad+ unlearning → Experiment artifacts: writes model and metrics
src/unlearn_neggrad_plus.py - CIFAR-10 → Original evaluation: provides evaluation data
src/evaluate_original.py - CIFAR-10 → Unlearned evaluation: provides evaluation data
src/evaluate_unlearned.py - CIFAR-10 → Reference evaluation: provides evaluation data
src/evaluate_retrained.py - Original evaluation → Experiment artifacts: writes evaluation
src/evaluate_original.py - Unlearned evaluation → Experiment artifacts: writes evaluation
src/evaluate_unlearned.py - Reference evaluation → Experiment artifacts: writes evaluation
src/evaluate_retrained.py - Experiment artifacts → Behavior comparison: supplies model outputs
- Behavior comparison → Experiment artifacts: writes comparison
src/compare_models.py - Experiment artifacts → Result summary: supplies results
src/summarize_results.py - Experiment artifacts → Trade-off plots: supplies plotted values
- Result summary → Researcher: reports findings
README.md - Trade-off plots → Researcher: visualizes trade-offs
README.md