Uncertainty Sets for Image Classifiers using Conformal Prediction
This paper introduces RAPS, a conformal prediction method with regularization, significantly reducing prediction set size while guaranteeing coverage.
Key Findings
Methodology
The approach extends conformal prediction by integrating a regularization term into the predictive set construction, which stabilizes tail probability estimates. It involves: • obtaining class probabilities from pretrained models; • ordering classes and applying a regularization to suppress tail noise; • calibrating the regularized sets via data splitting to select thresholds ensuring finite-sample coverage. The core innovation is the regularization of small, unreliable tail probabilities, combined with randomized mechanisms, to produce stable, compact predictive sets with formal guarantees across models and datasets. Experiments on Imagenet and Imagenet-V2 with ResNet-152 demonstrate that RAPS achieves coverage with sets 5-10 times smaller than baseline calibration methods.
Key Results
- On Imagenet, RAPS at 90% confidence level yields an average set size of 2.11, outperforming APS (19) and naive calibration, with coverage close to the target. Similar results hold for Imagenet-V2, confirming robustness across distributions.
- Adjusting the regularization parameter λ allows dynamic trade-offs between set size and adaptiveness, enabling smaller sets for easy samples and larger ones for difficult cases, while maintaining coverage.
- Across multiple models, RAPS consistently produces smaller, more stable prediction sets, with coverage accuracy exceeding 99%, validating its practical effectiveness.
Significance
This work advances uncertainty quantification in deep image classification by providing a theoretically grounded, scalable method that produces reliable, compact prediction sets. Its ability to guarantee coverage while reducing set size addresses a critical bottleneck in deploying AI in high-stakes environments like healthcare, autonomous driving, and security. The integration of regularization into conformal prediction offers a new paradigm for stabilizing tail probability estimates, making the approach suitable for real-world, noisy data scenarios. The method’s model-agnostic nature and minimal computational overhead facilitate broad adoption, promising to enhance trustworthiness and interpretability of AI systems.
Technical Contribution
The paper's main technical contribution is the development of RAPS, which incorporates a regularization term into the conformal prediction framework, controlling tail class probabilities. This leads to smaller, more stable predictive sets with formal finite-sample coverage guarantees. Theoretical analysis proves that, with appropriate parameter tuning, RAPS dominates fixed-size predictive sets and outperforms traditional calibration methods. The approach seamlessly integrates with existing deep classifiers, providing a practical, mathematically rigorous tool for uncertainty quantification in high-dimensional, noisy settings.
Novelty
This is the first systematic integration of regularization into conformal prediction for deep image classifiers, addressing tail probability instability. Unlike prior work that relies solely on probability calibration or ensemble methods, RAPS adaptively shrinks predictive sets by penalizing tail classes, achieving a balance between coverage and compactness. Its theoretical guarantees and empirical performance mark a significant step forward in reliable uncertainty quantification, especially in large-scale, high-dimensional vision tasks.
Limitations
- The method depends on selecting regularization parameters, which may require cross-validation or heuristic tuning, potentially affecting robustness in some scenarios.
- Extreme data shifts or highly noisy probability estimates can still challenge coverage guarantees, necessitating further robustness enhancements.
- While computationally efficient for standard models, scaling to extremely large architectures or real-time applications may pose efficiency challenges.
Future Work
Future research will focus on automating parameter tuning, extending the framework to multi-modal and sequential data, and integrating Bayesian or ensemble techniques for even more robust uncertainty estimates. Additionally, exploring online calibration and adaptive regularization strategies could further improve performance in dynamic environments. Broader applications in medical diagnostics, autonomous systems, and security will be pursued to validate and refine the approach in real-world settings.
AI Executive Summary
Deep convolutional neural networks have revolutionized image recognition, yet their uncertainty quantification remains a critical challenge. Existing calibration methods like Platt scaling improve probability estimates but lack formal guarantees, often resulting in overly large or unreliable predictive sets. This impairs their deployment in high-stakes domains such as medical diagnosis or autonomous driving, where understanding model confidence is vital.
The paper introduces RAPS, a conformal prediction framework enhanced with regularization, to address this gap. By penalizing tail class probabilities, RAPS stabilizes the estimation of unlikely classes, leading to smaller, more reliable predictive sets. The method involves: • extracting class probabilities from pretrained models; • ordering classes and applying a regularization term to suppress tail noise; • calibrating thresholds via data splitting to ensure finite-sample coverage guarantees.
Experiments on Imagenet and Imagenet-V2 with models like ResNet-152 demonstrate that RAPS achieves near-exact coverage with prediction sets 5-10 times smaller than traditional calibration approaches. The regularization parameter λ controls the adaptiveness, allowing the sets to shrink for easy samples while expanding for difficult ones, thus balancing size and coverage.
This work significantly advances the field of uncertainty quantification by providing a scalable, theoretically grounded method that enhances trustworthiness of deep classifiers. Its model-agnostic nature and minimal computational overhead facilitate broad adoption, promising to improve safety and interpretability in AI applications. Future directions include automatic parameter tuning, multi-modal extensions, and real-time deployment in dynamic environments.
Deep Analysis
Background
深度学习在图像识别中的应用不断扩大，模型的高准确率伴随对不确定性表达的需求日益增长。早期方法如贝叶斯神经网络和集成模型虽能提供一定的置信估计，但计算成本高、难以扩展。近年来，校准技术如Platt校准成为主流，但其在尾部概率校准上的不足导致预测集过大或偏离预期覆盖率。符合预测（Vovk et al., 2005）作为一种统计学工具，为模型提供严格的覆盖保证，逐渐引起关注。相关研究如 Romano et al. (2020)提出了适用于深度学习的符合预测算法，强调在保证覆盖的同时优化预测集大小。然而，尾部概率的不稳定性仍是挑战，限制了其在实际场景中的应用。本文在此基础上，结合正则化机制，提出更稳定、更紧凑的预测集方案，推动符合预测在图像分类中的落地。
Core Problem
深度图像分类模型的概率输出存在校准偏差，尤其在尾部类别估计中表现不佳，导致预测集过大，影响模型的实用性。现有校准方法如Platt校准未能提供严格的有限样本保证，且在尾部估计不稳定时表现欠佳。如何在保证覆盖的同时，缩小预测集，提升模型的可信度，成为亟待解决的问题。此外，模型尾部概率的噪声影响尾部类别排序，导致不可靠的预测集大小和内容，限制了符合预测的实际应用。解决这一问题，不仅需要理论上的覆盖保证，还需在实际中实现预测集的紧凑性和自适应性，满足高风险场景的需求。
Innovation
本论文的创新点主要在于：•引入正则化机制调节尾部类别概率，增强尾部估计的稳定性，减少噪声影响；•设计结合数据划分的符合预测算法，保证在任何模型和数据集上都能提供形式化的覆盖保证；•通过调节正则化参数，实现预测集的自适应调节，兼顾覆盖率与紧凑性，满足不同样本难易程度的需求。这些创新突破了传统校准和符合预测的局限，为深度学习模型的不确定性表达提供了更可靠的数学工具。
Methodology
- ��利用预训练模型输出类别概率，进行排序。•引入正则化项，抑制尾部类别的噪声影响，构建正则化预测集。•采用数据划分的符合预测策略，选择最小阈值保证覆盖。•在校准集上调节阈值，确保未来样本的覆盖率。•结合随机化机制，避免尾部类别排序的随机性影响，提升预测集稳定性。•通过理论分析，证明正则化参数调节下的覆盖保证与最优性。•在Imagenet系列数据上验证，预测集大小显著优于传统方法。
Experiments
采用Imagenet和Imagenet-V2数据集，使用ResNet-152等预训练模型。校准后，分别应用Naive、APS和RAPS方法，比较覆盖率和预测集大小。通过多次随机抽样，统计平均性能指标。调节正则化参数λ，观察预测集的变化趋势。评估指标包括：覆盖率、平均预测集大小、模型在不同难度样本上的表现。还进行了参数敏感性分析和不同模型的适应性测试，验证算法的稳健性和实用性。
Results
RAPS在所有模型和数据集上均实现了预期的覆盖率，平均预测集大小远小于APS和传统校准方案。例如，ResNet-152模型在90%置信水平下，RAPS预测集平均大小为2.11，远低于APS的19和单独校准的多倍。不同模型中，RAPS均表现出优越的紧凑性和稳定性，验证了其在实际应用中的潜力。调节λ后，预测集大小可在覆盖率和自适应性之间实现平衡，满足不同场景需求。
Applications
该方法适用于医疗影像诊断、自动驾驶、安防监控等高风险场景，为模型提供可靠的不确定性表达。用户只需在已有模型基础上进行校准，即可获得形式化保证的预测集，提升模型可信度。未来可结合在线学习和多模态信息，扩展到更复杂的应用场景，推动深度学习模型的安全部署。
Limitations & Outlook
目前算法依赖于概率输出的校准参数，参数调节可能影响性能。极端噪声或数据偏移可能削弱覆盖保证。计算成本虽低，但在超大模型和高维特征空间中仍存在效率挑战。未来需优化参数自动调节机制，增强鲁棒性。
Plain Language Accessible to non-experts
想象你在厨房做饭，手边有很多食材。每次做菜前，你会根据经验估算每种食材的用量，但有时候估算不准，可能多放或少放。为了确保菜的味道，厨师会提前准备一些备用方案，比如多准备一些调料或备用食材。这个过程就像模型在预测时输出的概率，但有时候这些概率不太准确，尤其是在尾部类别（少见的食材）上。为了避免做出不合适的菜，厨师会用一种特别的方法，把不确定的部分调节得更合理，确保最终菜的味道既不偏离，又不浪费食材。这个方法就像本文提出的正则化符合预测，既保证菜的质量，又节省材料，让厨房操作更高效、更可靠。
ELI14 Explained like you're 14
想象你在玩一个猜谜游戏，你要猜一个图片里是什么东西。通常，电脑会告诉你它猜的是什么，比如“猫”或者“狗”，并给出一个概率，比如“我觉得是猫的可能性是80%”。但有时候，这个概率不太准，尤其是那些不常见的动物或物品。为了更靠谱，科学家们设计了一种方法，让电脑不仅告诉你最可能的答案，还会给你一个“安全范围”，保证这个范围里一定有正确答案。这个方法会让电脑在猜的时候，避免只依赖不准的概率，而是用一种特别的技巧，把不确定的部分调节得更合理。这样一来，无论模型多不确定，它都能给出一个既小又可靠的答案范围，让你更有信心知道答案到底是什么。
Abstract
Convolutional image classifiers can achieve high predictive accuracy, but quantifying their uncertainty remains an unresolved challenge, hindering their deployment in consequential settings. Existing uncertainty quantification techniques, such as Platt scaling, attempt to calibrate the network's probability estimates, but they do not have formal guarantees. We present an algorithm that modifies any classifier to output a predictive set containing the true label with a user-specified probability, such as 90%. The algorithm is simple and fast like Platt scaling, but provides a formal finite-sample coverage guarantee for every model and dataset. Our method modifies an existing conformal prediction algorithm to give more stable predictive sets by regularizing the small scores of unlikely classes after Platt scaling. In experiments on both Imagenet and Imagenet-V2 with ResNet-152 and other classifiers, our scheme outperforms existing approaches, achieving coverage with sets that are often factors of 5 to 10 smaller than a stand-alone Platt scaling baseline.