The Miscalibration Problem
Modern deep learning models are notoriously "miscalibrated." If a model predicts a 90% probability that an image is a cat, its actual accuracy is often much lower—say, 70%. This over-confidence is a major hurdle for deploying AI in high-stakes environments like medicine or autonomous driving.
Until now, the research community has largely blamed over-parametrization. The prevailing wisdom suggests that because modern neural networks have millions of parameters, they can easily "memorize" training data, pushing their confidence scores to extreme, unwarranted highs. But is over-parametrization the only culprit?
This paper flips the script. By examining the simplest possible case—binary logistic regression where the data perfectly follows the model's assumptions—the authors investigate whether over-confidence is actually "baked in" to the underlying mathematics.
Theoretical Framework
To isolate the root cause of miscalibration, the authors employ a rigorous theoretical framework based on high-dimensional proportional limits. They analyze a scenario where both the number of samples ($n$) and the number of features ($d$) grow to infinity, but their ratio ($d/n$) remains a constant ($\kappa$).
The study focuses on unregularized Empirical Risk Minimization (ERM) for Generalized Linear Models (GLMs). By defining a precise metric for calibration error—the gap between a predicted probability and the actual conditional probability of the label—they use recent advances in high-dimensional statistics to derive a closed-form expression for the limiting calibration error. This allows them to determine exactly when and why a model becomes miscalibrated.
Key Insights: It's All About Curvature
The most striking finding is that logistic regression is inherently over-confident. Even when you have plenty of data ($n \gg d$) and the model is perfectly "correct" for the underlying data distribution, logistic regression still overestimates its accuracy.
The authors prove that this phenomenon is fundamentally tied to the curvature of the activation function. Any symmetric activation function that is concave on its positive half (like the standard sigmoid function used in logistic regression) will mathematically guarantee an over-confident classifier.
Surprisingly, this means over-confidence is not a universal rule. The authors demonstrate that by designing a custom activation function that is convex at certain points on the positive part, they can create a model that is actually under-confident—less sure of itself than it mathematically should be.
Furthermore, the "amount" of over-confidence scales with the ratio of parameters to samples ($d/n$, or $\kappa$). As you add more data relative to the number of features, the model becomes better calibrated, but the direction of the error (over-confidence) remains structurally the same.
Validation and Results
The paper validates its theoretical claims using a mix of synthetic and real-world data:
- Simulated Data: Using Gaussian input data generated from a "realizable" model, the analytical curves derived from their theory matched the simulated results almost perfectly, even at relatively small sample sizes ($d=100$).
- Real-World Data (CIFAR-10): To test the theory on real images, the authors used pseudo-labels generated by a teacher model to ensure the labels were perfectly consistent with a linear softmax model. The model remained over-confident, suggesting that the inherent bias of the logistic/softmax loss persists even when the strict Gaussian assumptions of the theory are relaxed.
Limitations and Future Frontiers
While groundbreaking, the formal mathematical proofs rely heavily on the assumption that input features follow a Gaussian distribution. Although empirical tests on CIFAR-10 suggest the findings are robust, a generalized theory for non-Gaussian data remains an open challenge.
Additionally, this study focuses on linear models (GLMs). While these are the foundational building blocks of neural networks, modern deep learning relies on complex, multi-layer feature learning. Extending this curvature analysis to nonlinear architectures is the next major frontier. Finally, because the paper focuses on unregularized models to isolate the effect of the loss function, future work will need to explore the interplay between inherent loss function bias and practical calibration techniques like $L_2$ regularization or weight decay.
Summary updated