📊 statistics

Singular model selection for trustworthy label-free classifier evaluation

This paper demonstrates that label-free classifier evaluation is a singular statistical problem where classical model selection criteria fail due to degenerate Fisher information, and proposes using singular and widely applicable Bayesian information criteria to accurately recover latent structures and ensure trustworthy performance estimates without ground-truth labels.

Original authors: Vincent Looten

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vincent Looten

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of artificial intelligence, we often trust machines to make decisions, but we rarely know how to trust the machines that judge those decisions. When a new computer program is built to recognize diseases in medical scans or identify birds in photographs, its performance is usually measured by comparing its answers to a known "gold standard"—a set of correct answers provided by human experts. But what happens when no such perfect answers exist? This is a common reality in fields like medicine, where a diagnosis might be uncertain, or in large-scale data projects where human experts are too expensive or too slow to label every single item. In these situations, researchers rely on a clever workaround: they ask several imperfect judges to look at the same items and compare their agreements. If a group of doctors or a crowd of workers mostly agree on a diagnosis, we assume they are likely right. However, this method hides a dangerous trap. The mathematical tools used to decide how many different types of answers exist in the data are often built on assumptions that break down exactly when the data is most difficult to interpret. When the signal is weak or the condition is rare, these standard tools can silently erase the complexity of the truth, collapsing distinct groups of answers into a single, misleading average.

A researcher has now shown that this collapse is not just a minor error in calculation, but a fundamental flaw in how we count the possibilities. They studied a specific type of statistical model used to evaluate these imperfect judges, known as a latent class model. In this setup, the "true" category of an item is hidden, and we try to guess it based on the noisy, sometimes conflicting votes of several raters. The researcher discovered that the geometry of this problem is "singular," a technical way of saying that the landscape of possible answers has a flat, degenerate spot where the standard rules of statistics no longer apply. In this singular zone, the usual methods for choosing the right number of hidden categories systematically fail. They tend to choose too few categories, merging distinct groups of data into one. This is not a matter of the computer being slightly off; it is a structural failure where the model decides that two different realities are actually the same thing.

To prove this, the researcher ran thousands of computer simulations where they knew the exact truth beforehand. They created scenarios where there were clearly two different types of items, but the signals from the raters were weak or the items were rare. When they applied the standard, widely used method known as the Bayesian Information Criterion, it almost always failed to see the second group. Instead, it reported that there was only one type of item, effectively erasing the distinction. This happened even when the researcher had a large amount of data. The standard method was so eager to keep the model simple that it ignored the evidence of a second group entirely. In contrast, the researcher tested two newer methods designed specifically for these tricky, singular situations. These new methods, which use a different way of measuring complexity, successfully found the second group in the vast majority of cases. They did not just find it; they did so without inventing fake groups when the truth was simple, a problem that plagues other, more lenient methods.

The consequences of getting this count wrong are severe. In a medical context, if a model collapses two distinct groups of patients into one, it cannot report accurate sensitivity or specificity—the measures of how well a test detects a disease versus how well it avoids false alarms. The researcher showed that when the standard method forces a collapse, the reported accuracy of the tests degenerates into a meaningless number that simply reflects the overall rate of the disease in the population. It becomes impossible to tell if a test is excellent or useless. To demonstrate this with real data, the researcher re-analyzed a famous dataset of skin cancer slides graded by seven pathologists, none of whom had a gold standard to compare against. When they used the standard method on smaller subsets of the data, it consistently missed a third, distinct group of slides where the doctors were confused or disagreed. The newer, singular-aware methods found this confused group immediately. This third group represented a real, difficult stratum of cases that the standard method had silently merged into the clear "yes" and "no" categories.

The researcher also tested this approach on a machine-learning dataset where crowd workers labeled images of birds. Here too, the standard method insisted there were only two types of images, while the new methods revealed a third, contested group of images that were genuinely harder to classify. By using the new methods, the researcher could isolate these difficult cases and show that they were not just noise, but a real, identifiable part of the data. The study confirms that the choice of how to count these hidden groups is not a minor technical detail but a prerequisite for trust. If the counting method is flawed, the entire evaluation of the machine learning system is built on a collapsed foundation. The researcher concludes that for any evaluation without a perfect reference standard, we must stop using the old, standard tools and adopt these new, singular-aware methods. Only then can we be sure that the performance numbers we report are honest, and that we are not mistaking a complex reality for a simple one.

Technical Summary: Singular Model Selection for Trustworthy Label-Free Classifier Evaluation

1. Problem Statement

The paper addresses the critical issue of trustworthy label-free evaluation of machine learning classifiers. In scenarios where ground-truth labels are unavailable (e.g., due to cost, unreliability, or the use of Large Language Models as raters), performance is estimated using Latent Class Models (LCMs). These models infer the accuracy of imperfect raters and the prevalence of latent classes based solely on the agreement patterns among raters.

The core problem identified is that standard model selection criteria, specifically the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), are asymptotically invalid for this task. LCMs are singular models by construction:

When latent classes have similar response profiles (common in weak signal or low prevalence regimes), the Fisher information matrix becomes degenerate.
The set of maximum likelihood estimates is not an isolated point but a curve or surface.
Consequently, the classical justification for BIC (which assumes a regular model with non-degenerate Fisher information and a penalty of $d/2 \log n$ , where $d$ is the parameter count) fails.
Using the Euclidean parameter count ( $d/2$ ) leads to systematic under-selection of the number of latent classes. In practice, this collapses a true multi-class structure into a single class, rendering reported sensitivities, specificities, and prevalence rates uninterpretable (collapsing to marginal base rates).

2. Methodology

The author proposes replacing standard criteria with methods derived from Singular Learning Theory (SLT), which accounts for the geometric degeneracy of the model.

Theoretical Framework

Real Log-Canonical Threshold (RLCT): For singular models, the asymptotic behavior of the marginal likelihood is governed by the learning coefficient $\lambda$ (where $\lambda \le d/2$ ) rather than $d/2$ . The log marginal likelihood behaves as $\hat{\ell} - \lambda \log n + (m-1)\log \log n$ .
Proposed Criteria:
1. Singular BIC (sBIC): Proposed by Drton and Plummer, this criterion replaces the $d/2$ penalty with bounds on the learning coefficient $\lambda$ derived from the model's geometry. It solves a fixed-point equation over the poset of nested models.
2. Widely Applicable BIC (WBIC): Proposed by Watanabe, this estimates the free energy (log marginal likelihood) directly via thermodynamic integration using a tempered posterior at inverse temperature $\beta = 1/\log n$ , without requiring a closed-form $\lambda$ .

Experimental Design

The study validates these methods through:

Simulation: A comprehensive simulation study varying signal strength ( $s$ ), sample size ( $n$ ), prevalence ( $\eta$ ), and number of tests/raters ( $p$ ). The ground truth includes $K^*=1$ (reverse check) and $K^*=2$ (primary arm), with extensions to $K^*=3$ and conditional dependence.
Real-World Applications:
1. Clinical: Re-analysis of a dataset with 118 histological slides graded by 7 pathologists (no gold standard).
2. Crowdsourced ML: Analysis of the "Bluebirds" dataset (108 images, 39 raters) where gold labels exist only for external validation, not fitting.
Validation:
- Thermodynamic Integration (TI): Direct computation of free energy to verify the effective learning coefficient is indeed lower than $d/2$ .
- Reverse Checks: Ensuring criteria do not over-select when the truth is a single class.
- Implementation: A reproducible Python engine validated against poLCA and sBIC packages.

3. Key Results

Performance in Simulation

BIC Failure: In weak signal and low prevalence regimes, standard BIC systematically under-selects, collapsing a true two-class structure to one class with high probability (e.g., $\Psi_{BIC} \le 0.02$ for low signal).
sBIC and WBIC Recovery: The singular criteria recover the correct structure ( $K=2$ ) across a broad transition band where BIC fails. For example, at $s=0.30, n=400$ , sBIC achieves a correct selection rate of 0.81 compared to BIC's 0.21.
AIC Comparison: While AIC avoids under-selection, it suffers from severe over-selection, inventing spurious classes (e.g., selecting $K=3$ when $K^*=2$ ) with rates that grow with sample size.
Mechanism: The recovery advantage is not generic statistical power but a precise correction of the penalty term. The empirical switchover points align with the theoretical locus where accumulated information ( $nD$ ) meets the specific penalty ( $\lambda \log n$ vs. $d/2 \log n$ ).
Conditional Dependence: The paper notes a critical distinction: selecting the number of classes is a singular problem (where sBIC helps), but modeling conditional dependence is a regular extension. If dependence is unmodeled, the singular penalty can lead to over-selection (absorbing dependence into a spurious class). Thus, the recommended workflow is to model dependence first, then select $K$ using a singular criterion.

Real-World Applications

Carcinoma Grading: On the full dataset ( $n=118$ ), all criteria agreed on $K=3$ . However, in subsampled experiments (smaller $n$ ), BIC under-selected to $K=2$ in 75% of cases at $n=80$ , whereas sBIC recovered $K=3$ in 97% of cases. The $K=3$ model revealed a clinically meaningful "discordant" stratum that BIC's $K=2$ model merged into the poles.
Crowdsourced Data: BIC consistently selected $K=2$ regardless of the number of raters. sBIC and WBIC recovered additional structure (up to $K=4$ ) as the number of raters increased, isolating a "contested" stratum validated by external gold labels.

4. Key Contributions

Operationalizing Singularity: The paper demonstrates that label-free evaluation is inherently a singular model selection problem and that standard criteria (BIC/AIC) are theoretically unjustified in the operating regimes of screening and weak signals.
Restoring Trustworthiness: It shows that using singular criteria (sBIC, WBIC) prevents the "collapse" of latent structures, thereby ensuring that reported performance metrics (sensitivity, specificity) are interpretable and not merely artifacts of a merged model.
Quantitative Mechanism: The study provides a quantitative explanation for selection failures, mapping the empirical recovery boundaries to the gap between the Euclidean penalty ( $d/2$ ) and the true learning coefficient ( $\lambda$ ).
Workflow Definition: It delineates a two-step workflow: (1) specify conditional dependence (a regular problem), and (2) select the number of latent classes using a singular criterion.
Validation: It closes the "circularity objection" regarding learning coefficients by using thermodynamic integration to confirm that the effective coefficient is indeed lower than the BIC bound in the singular regime.

5. Significance and Claims

The paper claims that honest complexity control is a precondition for trustworthy label-free evaluation.

Upstream Failure: The failure to select the correct number of classes is an upstream error that renders downstream accuracy estimates uninterpretable. When BIC collapses a two-class model to one, the reported sensitivity and specificity degenerate to the marginal base rate, and the Youden index drops to zero.
Generalizability: While demonstrated on diagnostic LCMs, the author argues this generative structure underlies recent approaches to evaluating Large Language Models (LLMs) and ensemble methods without ground truth.
Modest Scope: The author is modest about the limitations, noting that singular criteria rely on learning coefficient bounds (though verified as conservative) and depend on priors. They emphasize that the recovery guarantee holds conditional on the dependence structure being correctly specified. The paper does not claim to solve all LCM issues (like identifiability in very small samples) but specifically targets the model selection penalty in singular regimes.

In conclusion, the paper argues that the "trustworthiness" of AI evaluation metrics depends on using the correct asymptotic machinery (Singular Learning Theory) to avoid geometric failures that standard statistics silently overlook.