Freeze the network
Weights stay at initialization
The underlying randomly initialized weights are held fixed. Training changes score variables that determine a binary mask, so accuracy before weight training measures strong-ticket extraction.
Paper project · Sparse machine learning
Double-Scoring: Reliable Extraction of Strong Lottery Tickets
Bryce A. Christopherson · Jack Baretz · Darian Colgrove · Salah Dandan
University of North Dakota · 2026 preprint
Overview
The extraction problem, reparameterized.
Strong lottery tickets are sparse subnetworks that are already useful while their random weights remain untrained. Edge-popup searches for them by training scores over frozen weights, but it requires layerwise sparsity levels in advance. Double-scoring replaces that brittle search with competition in an augmented score space. The weight tensors do not grow; only the space used to decide which weights survive is enlarged.
Double the scores
Original and auxiliary coordinates
Each original score tensor competes with an equally shaped auxiliary tensor. A fixed-density top-k mask is chosen in the combined score space and then restricted back to the original coordinates.
Let sparsity emerge
No layerwise density tuple
Auxiliary coordinates act as a competitive reservoir. Their presence lets optimization induce the effective original-coordinate sparsity instead of requiring a separate density choice for every layer.
Interactive visualization
Watch a sparse subnetwork emerge.
Play the score-training run or scrub through it yourself, rotate the network, and select a handwritten digit to inspect how the emerging ticket processes it.
Citation
Use or reference this work.
The preprint is available on arXiv and in the University of North Dakota Scholarly Commons. The citation below uses the arXiv record.
BibTeX citation
@article{christopherson2026double,
title = {Double-Scoring: Reliable Extraction of Strong Lottery Tickets},
author = {Christopherson, Bryce A. and Baretz, Jack and Colgrove, Darian and Dandan, Salah},
year = {2026},
eprint = {2607.20555},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
doi = {10.48550/arXiv.2607.20555},
url = {https://arxiv.org/abs/2607.20555}
}