Fisher-Glass Learning: Classification Bounds and Subpopulation Audits

El Mustapha MansouriarXiv 2026 · arXiv

TL;DR

This paper introduces a framework to quantify "task information" that remains after filtering out background "nuisance" noise. By extending the Fisher-glass framework, the authors provide theoretical bounds on classification risk and practical auditing tools to identify subpopulations where a model might be secretly relying on spurious cues, even when its overall accuracy appears high.

What Problem Does This Paper Address?

Modern machine learning models are notorious for "cheating." A bird classifier might achieve 99% accuracy not because it recognizes feathers and beaks, but because it has learned that "water in the background equals waterbird." This is a spurious correlation or nuisance variation.

The core problem is that standard predictive metrics (like accuracy or log-likelihood) don't tell you why a model is right. If the environment changes—say, a waterbird is seen on a forest background—the model fails catastrophically. This paper addresses the gap between predictive performance and actual task-relevant information. It asks: How much of the model's decision is actually based on the task itself versus admissible nuisance noise?

Figure 1: Background sensitivity despite high control accuracy. (A,C) A matched Waterbirds image pair preserves the extracted bird foreground and label while changing background type. For the selected waterbird/water group, mean accuracy across five training seeds falls from 93.2% with same-type backgrounds to 40.6% with opposite-type backgrounds, a drop of 52.6 percentage points. (B) Schematic score geometry: projection removes the nuisance-aligned component of the task score, leaving a small residual.

Methods

The authors extend the Fisher-glass framework to classification. The key technical ideas include:

  • Efficient-Score Projection: They use a geometric approach to "project away" the components of the model's task score that can be explained by nuisance variables. What remains is the "surviving task score," which represents the information uniquely attributable to the task.
  • Local Minimax Bounds: They derive theoretical lower bounds on classification risk. Essentially, they prove that if the task information is heavily aligned with (and thus "collapsed" into) the nuisance span, the model will inevitably struggle to generalize when those nuisances change.
  • Population Information Audits: The paper provides a statistical "certificate" of information. This audit accounts for real-world issues like estimation error and regularization (ridge bias), ensuring that the claimed task information is statistically supported by the data.
  • Hidden-Slice Discovery: They propose an algorithm to search for "hidden slices"—subpopulations that are accurately classified but have dangerously low task information, making them prime candidates for future failure.

Datasets

The paper evaluates its framework on:

  • Waterbirds: A classic benchmark for spurious correlations where birds (waterbirds vs. landbirds) are typically seen on backgrounds (water vs. land) that correlate with their labels.
  • CivilComments-WILDS: A toxicity detection dataset where mentions of certain identities can act as nuisance variables that lead to biased predictions.
  • Synthetic Data: Used to verify the "geometric tail laws" that describe how task information vanishes across different environments.

Key Insights and Ideas

  • Information and Accuracy Can Diverge: A model can be nearly perfect on a specific group while having almost zero "task information." This means the model is essentially "flying blind" on the actual task and relying entirely on the background.
  • The Geometry of Vulnerability: By viewing task and nuisance scores as vectors in a space, the authors show that vulnerability is a matter of alignment. If the task vector is almost parallel to the nuisance span, the "surviving" signal is too weak to be robust.
  • Tail Behavior: The paper identifies that weak task information often follows a power-law "tail." Understanding this tail allows researchers to certify an "information floor" for a certain percentage of the population.
  • Controlled Interventions are Essential: To truly diagnose a model, you need to intervene. The authors use "background swaps" (keeping the bird but changing the scenery) to prove that their information metrics correctly predict which groups will fail.

Key Results

  • The Waterbirds "Crash": Using their diagnostic, the authors identified a group with 93.2% accuracy under normal conditions. When they performed a controlled background swap (keeping the bird but changing the background type), the accuracy plummeted to 40.6%—a massive 52.6 percentage point drop.
  • Hidden Vulnerabilities: Their search identified a subpopulation with 98.8% accuracy but extremely low "projected predictive Fisher information." This highlights that high accuracy can mask a total lack of robust task-relevant learning.
  • Information-Guided Training: By adding a Fisher-Glass penalty to the training objective, they were able to double the information certificate compared to standard Empirical Risk Minimization (ERM), though they note that information gains don't always translate immediately to accuracy gains on the source data.
Figure 2: (A) Selection of eligible slices. (B) Accuracy drop in the selected group when backgrounds are swapped.

Limitations and Future Work

  • Declared Nuisance Models: The audit is only as good as your "nuisance model." You have to specify what the potential nuisances are (e.g., background, lighting, identity mentions). If you miss a nuisance, the audit won't catch it.
  • Representation Dependence: The results depend on the specific feature representation and coordinate scaling used. The authors note that their method does not yet establish "representation invariance."
  • Local vs. Global: The theoretical bounds are local. Extending these to more global, non-linear regimes remains a challenge.
  • Human-in-the-loop: Future work could involve using these audits to help human annotators identify which types of new data are most needed to "break" spurious correlations.

Summary updated

Abstract

Accurate classifiers can fail when background cues change without altering the correct labels. Predictive performance alone does not quantify how much task information remains after accounting for such nuisance variation. Here we extend the Fisher-glass framework to connect nuisance-adjusted information with local classification difficulty and subpopulation auditing. We derive lower bounds on local minimax excess classification risk under explicit distinguishability and decision-separation conditions, alongside statistical audits accounting for estimation error and regularization. On Waterbirds, a feature-based diagnostic of task-background alignment selects an accurately classified group, with selection frozen before held-out evaluation. Across five training seeds, background replacements preserving the extracted bird foreground and labels yield 93.2% accuracy when background type is preserved and 40.6% when it changes, a controlled difference of 52.6 percentage points. Separately, search using task labels and a fixed nuisance direction identifies an independently evaluated population with 98.8% membership-weighted accuracy and low estimated projected predictive Fisher information. Training and acquisition experiments show that information gains and predictive gains can diverge. These findings support joint assessment of predictive performance, information under declared nuisance models and controlled sensitivity when diagnosing vulnerability and evaluating model improvements.
BibTeX
@misc{mansouri2026fisher,
  author = {El Mustapha Mansouri},
  title = {{Fisher-Glass Learning: Classification Bounds and Subpopulation Audits}},
  year = {2026},
  eprint = {2611.12702},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.12702},
}
Figure