On the Duality between Feature and Sample Screening
arXiv:2609.02418
2026
Memory
1 ideas extracted · analyzed Sep 3, 2026
What the math gives to ML
The paper gives a structural link between feature pruning and sample pruning through Fenchel–Rockafellar duality. Its transferable asset is the equivariance identity: feature screening in the primal is exactly represented as sample screening after dualization, because matrix columns and rows exchange roles. A practical neural-network use is to apply this framework to convex subproblems such as a frozen-encoder linear probe, sparse adapter, or classifier, then transfer the discovered feature masks to channels or embedding coordinates and use dual-derived scores for data pruning.
Ideas from this paper
✗ Mechanism failed
2026
Train a convex surrogate attached to a neural network, such as a sparse linear probe or adapter, and use its primal and Fenchel-dual representations to produce mutually consistent feature and sample masks. The same screening operation can be evaluated on columns of the primal design matrix or rows of the dual representation, allowing data and feature reduction without independently tuned heuristics.
Useful7/10
Difficulty5/10
Novelty6/10