How naresh-gudipalli/few-shot-image-classification-can-intelligent-data-curation-replace-large-training-datasets- works
This repository documents controlled few-shot image-classification experiments: train a full-data SimpleCNN baseline, select small training subsets using different strategies, train on those subsets, and compare results across MNIST, CIFAR-10, seeds, and subset sizes. The notebooks are the product’s experiment stages. The README documents Phase 4’s use of Phase 1 checkpoints for embedding-based selection and Phase 5’s use of Phase 1 checkpoints for forgetting-event scoring. Source excerpts were unavailable, so the graph shows only these documented relationships.
Main entry points: phase1-baseline-final.ipynb Full-data baseline · phase-5-hard-examples-mining.ipynb Forgetting mining · phase4-k-means.ipynb K-means coreset · REPRODUCIBILITY.md Phase comparisons
Components 10
Baseline Training
- Full-data baseline
phase1-baseline/phase1-baseline-final.ipynb
Subset Selection
- Random sampling
phase2-random-subsampling/phase2_random_subsampling.ipynb - Class-balanced selection
phase3-class-balanced/phase3-class-balanced.ipynb - K-means coreset
phase4-k-means/phase4-k-means.ipynb
Hard Mining
- Forgetting mining
phase-5-hard-examples-mining via Forgetting examples/phase-5-hard-examples-mining.ipynb - Mining diagnostics
phase-5-hard-examples-mining via Forgetting examples/fpv-p5-hard-mining-examples.ipynb
Evaluation
- Phase comparisons
REPRODUCIBILITY.md
Other components
- Experimenter
- MNIST
- CIFAR-10
Connections 23
- Experimenter → Full-data baseline: runs experiment
- Experimenter → Random sampling: runs experiment
- Experimenter → Class-balanced selection: runs experiment
- Experimenter → K-means coreset: runs experiment
- Experimenter → Forgetting mining: runs experiment
- MNIST → Full-data baseline: trains on
- CIFAR-10 → Full-data baseline: trains on
- MNIST → Random sampling: samples from
- CIFAR-10 → Random sampling: samples from
- MNIST → Class-balanced selection: samples from
- CIFAR-10 → Class-balanced selection: samples from
- MNIST → K-means coreset: embeds images
- CIFAR-10 → K-means coreset: embeds images
- Full-data baseline → K-means coreset: provides checkpoints
- MNIST → Forgetting mining: scores examples
- CIFAR-10 → Forgetting mining: scores examples
- Full-data baseline → Forgetting mining: provides checkpoints
- Forgetting mining → Mining diagnostics: examines mining
- Full-data baseline → Phase comparisons: baseline results
- Random sampling → Phase comparisons: random results
- Class-balanced selection → Phase comparisons: balanced results
- K-means coreset → Phase comparisons: coreset results
- Forgetting mining → Phase comparisons: mining results