Abstract / TL;DR Existing models for Generalized Category Discovery (GCD) suffer from a "confidence bias" where entropy regularization—intended to prevent overfitting—indiscriminately suppresses certainty on all unlabeled data. To address this, researchers introduce Reliable Confidence Alignment (RCA), a framework leveraging Evidential Deep Learning to calibrate uncertainty. RCA bridges the gap between labeled and unlabeled samples, significantly boosting performance on known classes without sacrificing the ability to discover new ones.
The Challenge: "Unknown Unknowns" in Categorization
In real-world applications, AI models frequently encounter "unknown unknowns." Generalized Category Discovery (GCD) is the task of taking a model trained on a limited set of labeled data (known classes) and asking it to categorize a large pool of unlabeled data containing a mix of those known classes and entirely novel ones.
The core challenge in GCD is the stability-plasticity dilemma: how can a model remain stable enough to remember known classes while being "plastic" enough to cluster new ones? Current state-of-the-art methods rely on entropy regularization to prevent the model from becoming overconfident and collapsing all novel instances into known categories.
The Problem: Confidence Bias and the Distributional Wedge
While entropy regularization prevents class collapse, the authors identify a systemic flaw: Confidence Bias. Entropy regularization acts as a blunt instrument, indiscriminately suppressing predictive certainty for all unlabeled instances.
This suppression creates a "distributional wedge." To the model’s uncertainty estimator, a labeled image of a dog and an unlabeled image of the exact same dog look completely different. Because the model is penalized for being confident about unlabeled data, it is forced to misclassify familiar concepts as "new" clusters. This bias fundamentally forces models to sacrifice stability on known classes to achieve plasticity for new ones.
The Solution: Reliable Confidence Alignment (RCA)
To resolve this bias, the authors propose Reliable Confidence Alignment (RCA), a plug-and-play framework grounded in Evidential Deep Learning (EDL).
Standard Softmax functions only output point-estimate probabilities, indicating which class is most likely without capturing the nuance of epistemic uncertainty. RCA instead treats class probabilities as latent variables governed by a Dirichlet distribution. This allows the model to explicitly represent "evidence" and align the structure of uncertainty.
The RCA framework consists of two primary modules:
- Reliable Anchor for Certainty (RAC): This module uses the labeled dataset to establish "anchors" of high certainty. By optimizing a Variational Lower Bound (ELBO), RAC ensures the model maintains high-fidelity confidence on known categories, directly counteracting the suppression caused by entropy regularization.
- Cross-view Confidence Alignment (CCA): Grounded in PAC-Bayesian theory, CCA propagates this grounded reliability to the unlabeled discovery set. It forces the model to maintain consistency across different "views" (data augmentations) of the same image. Crucially, CCA aligns both the first-order (expected class probability) and second-order (fine-grained geometry of the distribution) information on the probability simplex.
Key Insights and Results
The authors evaluated RCA across coarse-grained benchmarks (CIFAR-10, CIFAR-100, ImageNet-100) and fine-grained benchmarks (CUB, Stanford Cars, FGVC-Aircraft), where subtle inter-class differences make confidence calibration critical.
Tested as an add-on to both "weak" (SimGCD) and "strong" (AFGCD) baselines, RCA demonstrated consistent and significant improvements:
- Correcting "False New" Errors: Empirical analysis reveals that RCA's primary benefit is a massive reduction in "False New" errors—instances where the model mistakenly labels a known class in the unlabeled pool as a new category due to artificially suppressed confidence.
- Performance Gains: On CIFAR-100, RCA improved overall accuracy by 3.5%, with a striking 4.8% jump in performance on known (Old) classes. On the fine-grained CUB dataset, Old-class performance was boosted by approximately 5%.
- Training Stability: Unlike previous methods where accuracy on known classes decays late in training as the model "discovers" new clusters, RCA maintains high stability throughout the entire training process.
Limitations and Future Directions
While RCA significantly improves the balance between known and novel classes, it introduces a slight trade-off. In some cases, there is a moderate increase in "False Old" errors (misclassifying a new class as a known one). However, this is heavily outweighed by the gains in overall accuracy and known-class retention.
The authors suggest that this "uncertainty-aware alignment" could be a powerful tool beyond GCD. By effectively calibrating epistemic uncertainty, RCA's principles could benefit fields like Open-Set Recognition, Continual Learning, and Safety-Critical Autonomous Systems, where distinguishing between "known-unknowns" and "unknown-unknowns" is paramount.
Summary updated