Model Calibration

Machine learning approaches for image classification have led to impressive advances in that field. For example, convolutional neural networks are able to achieve remarkable image classification accuracy across a wide range of applications in industry, defense, and other areas. While these machine learning models boast impressive accuracy, a related concern is how to assess and maintain calibration in the predictions these models make. A classification model is said to be well calibrated if its predicted probabilities correspond with the rates events actually occur. While there are many available methods to assess machine learning calibration and recalibrate faulty predictions, less effort has been spent on developing approaches that continually monitor predictive models for potential loss of calibration as time passes. We propose a cumulative sum-based approach with dynamic limits that enable detection of miscalibration in both traditional process monitoring and concept drift applications. This enables early detection of operational context changes that impact image classification performance in the field. The proposed chart can be used broadly in any situation where the user needs to monitor probability predictions over time for potential lapses in calibration. Importantly, our method operates on probability predictions and event outcomes and does not require under-the-hood access to the machine learning model.

The "alignment tax" of post-training is typically framed as a drop in task accuracy. We show it also involves a severe loss of calibration, making models overconfident, less reliable, and model outputs less diverse. We show that this trade-off can be navigated effectively via a simple post-hoc intervention: interpolating between a model's weights before and after alignment. Crucially, this is not a strict trade-off. We find that the process consistently reveals Pareto-optimal interpolations - models that improve accuracy beyond both parents while substantially recovering the calibration lost during alignment. Our work demonstrates that simple model merging provides a computationally efficient method for mitigating the full scope of the alignment tax, yielding models that are more capable and more reliable.

Accurate prediction of grape phenology is essential for timely vineyard management decisions, such as scheduling irrigation and fertilization, to maximize crop yield and quality. While traditional biophysical models calibrated on historical field data can be used for season-long predictions, they lack the precision required for fine-grained vineyard management. Deep learning methods are a compelling alternative but their performance is hindered by sparse phenology datasets, particularly at the cultivar level. We propose a hybrid modeling approach that combines multi-task learning with a recurrent neural network to parameterize a differentiable biophysical model. By using multi-task learning to predict the parameters of the biophysical model, our approach enables shared learning across cultivars while preserving biological structure, thereby improving the robustness and accuracy of predictions. Empirical evaluation using real-world and synthetic datasets demonstrates that our method significantly outperforms both conventional biophysical models and baseline deep learning approaches in predicting phenological stages, as well as other crop state variables such as cold-hardiness and wheat yield.

In many real-world applications, ensuring the robustness and stability of deep neural networks (DNNs) is crucial, particularly for image classification tasks that encounter various input perturbations. While Mixup-based data augmentation techniques have been widely adopted to enhance the resilience of trained models against such perturbations, our experiments reveal an important corruption robustness-calibration trade-off: stronger Mixup-based augmentation can improve robustness against corrupted data while substantially increasing expected calibration error (ECE). To address this challenge, we introduce DRO-Augment, a framework that integrates Wasserstein Distributionally Robust Optimization (W-DRO) with various Mixup-based data augmentation strategies to mitigate this trade-off. Our method substantially reduces ECE under strong Mixup-based augmentation while largely preserving corruption accuracy across CIFAR-10, CIFAR-100, CIFAR-10-C, and CIFAR-100-C. On the theoretical side, we establish novel generalization error bounds for neural networks trained using a variation-regularized loss function with augmented data, closely related to the W-DRO problem. Furthermore, we introduce a refined CIFAR-C benchmark that corrects inconsistencies in corruption intensities, providing a more reliable evaluation for future robustness research.

Deep neural networks have been increasingly used in safety-critical applications such as medical diagnosis and autonomous driving. However, many studies suggest that they are prone to being poorly calibrated and have a propensity for overconfidence, which may have disastrous consequences. In this paper, unlike standard training such as stochastic gradient descent, we show that the recently proposed sharpness-aware minimization (SAM) counteracts this tendency towards overconfidence. The theoretical analysis suggests that SAM allows us to learn models that are already well-calibrated by implicitly maximizing the entropy of the predictive distribution. Inspired by this finding, we further propose a variant of SAM, coined as CSAM, to ameliorate model calibration. Extensive experiments on various datasets, including ImageNet-1K, demonstrate the benefits of SAM in reducing calibration error. Meanwhile, CSAM performs even better than SAM and consistently achieves lower calibration error than other approaches

Foundation model ensembles can improve accuracy and robustness, but full fine-tuning and parameter-space merging are often expensive or impossible when candidate models have different architectures. We propose Optimizable Model Averaging (OMA), a black-box approach that learns nonnegative model-combination weights, one scalar coefficient per model, to average the candidate models' predictive probabilities, leaving all model parameters frozen, by minimizing regularized predictive entropy. A central challenge is that minimizing predictive entropy can favour confidently incorrect models: increased confidence need not imply improved accuracy or calibration. We therefore examine regularization as a mechanism for controlling adaptation and empirically evaluate its effect on calibration, observing lower expected calibration error in selected configurations, alongside better performance-calibration trade-offs than model output averaging. As foundation models continue to advance, these approaches will enable the incorporation of future, potentially much better models to improve performance on challenging text and image classification tasks.

Open-set recognition refers to the problem in which classes that were not seen during training appear at inference time. This requires the ability to identify instances of novel classes while maintaining discriminative capability for closed-set classification. OpenMax was the first deep neural network-based approach to address open-set recognition by calibrating the predictive scores of a standard closed-set classification network. In this paper we present MetaMax, a more effective post-processing technique that improves upon contemporary methods by directly modeling class activation vectors. MetaMax removes the need for computing class mean activation vectors (MAVs) and distances between a query image and a class MAV as required in OpenMax. Experimental results show that MetaMax outperforms OpenMax and is comparable in performance to other state-of-the-art approaches.

The construction of most supervised learning datasets revolves around collecting multiple labels for each instance, then aggregating the labels to form a type of "true" label. We question the wisdom of this pipeline by developing a (stylized) theoretical model of this process and analyzing its statistical consequences, showing how access to non-aggregated label information can make training well-calibrated models more feasible than it is with cleaned labels. The entire story, however, is subtle, and the contrasts between aggregated and fuller label information depend on the particulars of the problem, where estimators that use aggregated information exhibit robust but slower rates of convergence, while estimators that can effectively leverage all labels converge more quickly if they have fidelity to (or can learn) the true labeling process. The theory makes several predictions for real-world datasets, including when non-aggregate labels should improve learning performance, which we test to corroborate the validity of our predictions.