SING: Analyzing Semantic Invariants in Classifiers

CVPR 2026

Harel Yadid Harel Yadid
Meir Yossef Levi Meir Yossef Levi
Roy Betser Roy Betser
Guy Gilboa Guy Gilboa
Technion

Abstract

All classifiers, including state-of-the-art vision models, contain invariants induced by the geometry of their linear mappings. These invariants reside in the classifier null space and produce equivalent inputs with identical outputs.

SING (Semantic Interpretation of the Null-space Geometry) translates this hidden geometric behavior into human-interpretable semantics by mapping classifier features into a multimodal space and quantifying semantic drift. The framework supports single-image analysis and model-level comparisons.

SING teaser figure showing benign and problematic invariants

Method

Consists of four steps:

Extract Null/Principal Projectors

Method step 1

Decompose classifier-head weights with SVD to obtain principal and null subspaces.

Train a linear Translator

Method step 2

Map classifier features into a shared vision-language embedding space.

Create Equivalent Pair

Method step 3

Construct equivalent features by null space projection/manipulation while preserving logits.

Translate, Measure, Visualize

Method step 4

Quantify semantic drift and visualize induced changes across attributes/classes.

Key Insights

Semantic Tool for Invariants

Semantic tool for invariants

SING links classifier null-space geometry and the invariants it induces to human-readable semantic explanations through equivalent-pair analysis.

Model Comparison

Model comparison contribution

We compared architectures by quantifying semantic leakage into null directions, highlighting how well class semantics are preserved across invariant space.

Open-Vocabulary Class Analysis

Open-vocabulary class analysis contribution

The framework supports systematic class-level probing to reveal sensitivity to concepts and inspect semantic invariants.

Practical Diagnostics

Practical diagnostics contribution

Model-level & single-image workflows provide reproducible semantic diagnostics with both quantitative scores and qualitative visual evidence.

Angle-to-Visual Intuition

Angle-to-visual intuition animation

More Papers to Explore

BibTeX

@misc{yadid2026sing,
      title={Make it SING: Analyzing Semantic Invariants in Classifiers}, 
      author={Harel Yadid and Meir Yossef Levi and Roy Betser and Guy Gilboa},
      year={2026},
      eprint={2603.14610},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.14610}
}