google-deepmind/hope v1.0

HOPE v1.0: Initial Release

Repository: google-deepmind/hope

Tag: v1.0

Published: 2026-10-27T21:29:19Z

Prerelease: no

Release notes: Initial open-source release accompanying *Hilbert Operator for Progressive Encoding (HOPE): A Mathematical Framework for Deconstructing Learned Representations in Deep Networks* (Hossein Mobahi and Peter L. Bartlett, 2026).

Highlights

• Data-Free Progressive Model Compression (`hope.py`, `compress_resnet50.py`): Closed-form progressive compression (PRUNE, MERGE, EVICT) and lossless step-by-step decoding of pretrained ResNet-50 models using only weights and stored BatchNorm statistics.
• Structured Compression Baselines & Plotting (`hope_baselines.py`, `plot_compression_curves.py`): Reference implementations of seven data-free structured neuron compression baselines and trajectory visualization utilities.
• Dispersed Elastic Fine-Tuning (`deft.py`): Cross-domain transfer learning pipeline (CIFAR-100 to SVHN) partitioning pretrained networks into a frozen Universal Core and a plastic Slack via HOPE's progressive curriculum, together with transfer baselines (Linear Probe, Full Fine-Tuning, LoRA, Surgical Fine-Tuning, Magnitude Pruning + Fine-Tuning, and lottery-ticket controls).
• Self-Contained Test Suite (`test_hope.py`): Fast unit and integration tests verifying closed-form operators, determinism, and lossless progressive decoding.