The paper proposes replacing standard affine neural layers with a structured-linear-plus-residual decomposition to improve training stability and interpretability.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Structured Transformations for Stable and Interpretable Neural Computation
The paper proposes replacing standard affine neural layers with a structured-linear-plus-residual decomposition to improve training stability and interpretability.