Adaptive Confidence Regularization for Multimodal Failure Detection

AI summary

Overview

Research area: Multimodal machine learning, specifically failure detection (also called misclassification detection or selective classification) in multimodal models that combine video, optical flow, audio, image, and LiDAR inputs.

Technical level: Intermediate to Advanced. The paper assumes familiarity with softmax confidence scoring, out-of-distribution (OOD) detection metrics such as AUROC and FPR95, and standard multimodal fusion architectures.

Scope: The paper proposes Adaptive Confidence Regularization (ACR), a training framework with two components — an Adaptive Confidence Loss and Multimodal Feature Swapping — designed to make multimodal models better at flagging their own misclassifications.

What This Paper Is About

Multimodal models fuse inputs such as video and optical flow, and they often predict more accurately than unimodal ones. However, when they are wrong, they can still be confidently wrong, which is dangerous in safety-critical settings. The authors study why fused predictions sometimes end up less confident than one of their own unimodal branches (a pattern they call "confidence degradation"), and they build a training method that penalizes this behavior and synthesizes artificial failure examples so the model learns to recognize and reject unreliable predictions.

Key Contributions

1. Empirical case for multimodal failure detection: The authors show on the HMDB51 dataset that a simple fusion of video and optical flow improves failure detection metrics (AURC, AUROC, FPR95) over unimodal baselines, and that advanced OOD scores such as Energy, Entropy, and MaxLogit are outperformed by a simple Maximum Softmax Probability (MSP) baseline on failure detection — motivating methods built specifically for this problem.

2. Identification and formalization of confidence degradation: They define confidence degradation as the case where the fused multimodal prediction's confidence is strictly lower than that of at least one unimodal branch, and show across HMDB51, EPIC-Kitchens, HAC, and Kinetics-600 that misclassified samples exhibit this pattern far more often than correct ones.

3. The ACR framework: They introduce an Adaptive Confidence Loss that penalizes fused confidence falling below unimodal confidence, plus Multimodal Feature Swapping, a feature-space outlier synthesis technique that creates failure-aware training samples without external outlier datasets.

4. Extensive evaluation across datasets, modalities, and settings: Experiments cover four action recognition datasets with video and optical flow, additional modality combinations on HAC, distribution-shift corruptions, multimodal OOD detection, combined OOD-plus-misclassification detection, alternative backbones (I3D, TSN), and 3D semantic segmentation on SemanticKITTI. The authors report ACR outperforming prior best methods by up to 9.58% in AURC, 1.63% in AUROC, and 15.45% in FPR95. Code is stated to be released at https://github.com/mona4399/ACR.

Main Findings

• Confidence degradation tracks misclassification: Misclassified samples show markedly higher rates of confidence degradation than correct ones — increases of 32.4% on HMDB51, 23.1% on EPIC-Kitchens, 52.4% on HAC, and 22.0% on Kinetics-600.

• Gains on video + optical flow (Table 1): On HMDB51, ACR reaches AURC 19.97, AUROC 92.02, FPR95 41.96, ACC 87.23 versus MSP's 29.56, 88.28, 52.07, 86.20. On EPIC-Kitchens: 103.25, 79.27, 71.58, 75.20 (MSP: 115.03, 76.52, 76.84, 74.25). On HAC: 27.41, 91.48, 39.39, 84.86 (MSP: 42.90, 89.27, 66.67, 82.11). On Kinetics-600: 41.85, 88.99, 55.89, 81.45 (MSP: 46.29, 87.33, 61.29, 81.24). Accuracy improves on every evaluated dataset.

• Both components are complementary (Table 2, HMDB51): MSP 29.56 / 88.28 / 52.07 / 86.20; Adaptive Confidence Loss alone 24.48 / 90.32 / 43.97 / 86.77; Multimodal Feature Swapping alone 25.11 / 90.55 / 46.22 / 86.43; combined 19.97 / 92.02 / 41.96 / 87.23.

• Robust across modality combinations (Table 3, HAC): Averaged over video+audio, optical flow+audio, and video+optical flow+audio, ACR gives average improvements of 8.39% in AURC, 1.51% in AUROC, and 10.65% in FPR95 over the strongest baseline, and raises accuracy from 81.19% to 82.42%.

• Performance holds under distribution shift: With corruptions applied only to the video modality at severity level 5 (Defocus Blur, Frost, Brightness, Pixelate, JPEG Compression) on HAC, ACR shows substantially improved AURC in the majority of tested cases.

• Works with different backbones (Table 4, HMDB51): Using I3D and TSN instead of SlowFast, ACR obtains AURC 27.73, AUROC 90.00, FPR95 51.56, accuracy 85.40, better on AURC and FPR95 than MSP (31.77, 88.17, 58.59, 85.40) and the other listed baselines.

• Also improves multimodal OOD detection (Table 6, HMDB51 as ID): Averaged over Kinetics-600, UCF101, EPIC-Kitchens, and HAC, adding ACR on top of GEN gives FPR95 17.96 and AUROC 95.84 versus 23.24 and 94.49 for the A2D+NP-Mix (AN) strategy, at ID accuracy 87.23 versus 86.89.

• Handles OOD and misclassified samples together (Table 5): When OOD samples from HAC are mixed into HMDB51 test data, ACR achieves AUROC 96.82, FPR95 20.47, ACC 87.23, improvements of 1.47% in AUROC, 4.45% in FPR95, and 0.57% in ACC over the strongest baseline.

• Multimodal Feature Swapping beats other feature-space augmentations (Table 7, HMDB51): MFS 19.97 / 92.02 / 41.96 / 87.23 versus Random Noise 24.86 / 90.82 / 42.86 / 86.43, Random Drop 22.80 / 91.24 / 46.09 / 86.89, and Feature Mixing 21.79 / 91.33 / 42.11 / 87.00.

• Extends to 3D semantic segmentation (Table 8, SemanticKITTI): With image and LiDAR modalities, ACR yields AURC 21.90, AUROC 84.51, FPR95 52.51, mIoU 63.56, versus MSP's 33.90, 79.97, 55.49, 59.25.

• Visualization: Confidence-score distributions on HMDB51 are more clearly separated between correct and incorrect predictions under ACR than under MSP.

• Note on stated scope: The abstract and conclusion describe evaluation across four datasets and three modalities, while the introduction states "five datasets and five modalities." The tables cover five datasets (HMDB51, EPIC-Kitchens, HAC, Kinetics-600, SemanticKITTI) and five modalities (video, optical flow, audio, image, LiDAR), but the paper's own totals are stated inconsistently.

Methodology in Plain English

The authors start from a diagnostic observation rather than a new architecture. They train standard multimodal models with only cross-entropy loss and compare, for each sample, the fused prediction's confidence against the confidence of each single-modality branch. When the fused confidence is lower than at least one branch's confidence, they label that a case of "confidence degradation." They find this pattern is much more common among misclassified samples, which suggests it can be used as a training signal.

Their first component, the Adaptive Confidence Loss, is a penalty term: it adds zero cost when the fused confidence is at least as high as every unimodal confidence, and otherwise charges a penalty proportional to how far the fused confidence falls below the higher unimodal confidences. Minimizing this pushes the fusion to integrate complementary evidence rather than lose confidence, and also discourages one modality from being "confidently wrong."

Their second component, Multimodal Feature Swapping, creates artificial failure examples without any external outlier dataset. For a sample's feature embeddings, it randomly picks a number of feature dimensions from each modality and swaps them across modalities, then concatenates the modified embeddings. Since the swapped amount varies, the resulting synthetic examples can be near the real data (hard cases) or far away (obvious outliers). These synthetic examples are supervised with soft labels that interpolate between the true class label and an extra outlier class, with the interpolation weight set to the fraction of features swapped. Because only part of the representation is exchanged, the synthetic sample keeps within-modality semantics but breaks cross-modal consistency — mimicking a realistic failure mode.

The total training loss combines the ordinary classification loss, the cross-entropy loss on the synthetic outliers, and the Adaptive Confidence Loss weighted by a hyperparameter. At test time, the method simply uses the standard MSP score as the confidence measure and evaluates only on the known classes.

Experimental setup: MMAction2 toolkit; SlowFast (initialized from Kinetics-400) for video and for a slow-only optical flow encoder; ResNet-18 initialized from VGGSound for audio; Adam optimizer with learning rate 0.0001 and batch size 16; hyperparameters λ_acl = 2.0, n_min = 32, n_max = 256; 50 epochs on an NVIDIA RTX 3090, with model selection on the validation set. Baselines include MSP, MaxLogit, Energy, Entropy, DOCTOR, OpenMix, Mixup, RegMixup, CRL, and A2D. Metrics are AURC (scaled by 10^3), AUROC, FPR95 (false positive rate at 95% TPR), and accuracy on in-distribution data.

The appendix adds two theoretical results: adding modalities cannot increase the conditional entropy of the label given the input (via the data processing inequality), and lower predictive confidence corresponds to higher conditional entropy, which by Fano's inequality raises the minimum theoretical bound on error.

Why This Matters

Impact on research: Failure detection has been studied extensively for unimodal models, and OOD detection for multimodal models, but the paper argues the extension of failure detection to multimodal systems was largely unexplored. It provides evidence that OOD scoring functions transfer poorly to this task (MSP beats Energy, Entropy, and MaxLogit), formalizes a multimodal-specific failure signature (confidence degradation), and offers a baseline method plus code that future work can build on.

Real-world applications:

• Self-driving vehicles, where misclassified perception outputs can lead to collisions.
• Medical diagnostics, where an unreliable multimodal prediction should be routed to a human expert.
• Robotics and industrial automation in safety-critical environments.
• Video or audio analysis pipelines where corrupted or conflicting sensor streams should trigger abstention rather than a confident wrong answer.

Industry relevance: Systems deployed in regulated, high-stakes settings need a mechanism to say "I am not sure" rather than only maximizing accuracy. ACR is designed to be lightweight — outliers are generated in feature space, so no large external outlier dataset is required — and the paper reports that it improves accuracy while improving failure detection, which matters for practitioners who cannot trade away predictive performance for reliability.

Future Directions

• Adversarial robustness: The limitations section in the provided content is truncated mid-sentence at the discussion of ACR's behavior against "sophisticated adversarial" inputs, so the authors' specific conclusions there are not reported.
• Scaling modality counts and fusion designs: Evaluation covers up to three modalities for action recognition plus image+LiDAR for segmentation; how the Adaptive Confidence Loss behaves as the number of branches grows, and how it interacts with other fusion mechanisms, is an open question.
• Choosing the swap budget: Multimodal Feature Swapping relies on n_min, n_max, and λ_acl, and the paper shows outlier difficulty varies with the number of swapped dimensions; automating or adapting these settings is a natural next step.
• Beyond classification: The SemanticKITTI experiment tests 3D semantic segmentation, but the framework's benefit for other dense-prediction or generative multimodal tasks is not established.

Target Audience

Researchers and practitioners working on uncertainty estimation, selective prediction, OOD detection, or trustworthy multimodal machine learning. It will be most useful to engineers building safety-critical multimodal systems who need a practical training-time method for flagging unreliable predictions, and to academics looking for a well-characterized, quantitatively supported failure signature in multimodal fusion.

Authors’ abstract

The deployment of multimodal models in high-stakes domains, such as self-driving vehicles and medical diagnostics, demands not only strong predictive performance but also reliable mechanisms for detecting failures. In this work, we address the largely unexplored problem of failure detection in multimodal contexts. We propose Adaptive Confidence Regularization (ACR), a novel framework specifically designed to detect multimodal failures. Our approach is driven by a key observation: in most failure cases, the confidence of the multimodal prediction is significantly lower than that of at least one unimodal branch, a phenomenon we term confidence degradation. To mitigate this, we introduce an Adaptive Confidence Loss that penalizes such degradations during training. In addition, we propose Multimodal Feature Swapping, a novel outlier synthesis technique that generates challenging, failure-aware training examples. By training with these synthetic failures, ACR learns to more effectively recognize and reject uncertain predictions, thereby improving overall reliability. Extensive experiments across four datasets, three modalities, and multiple evaluation settings demonstrate that ACR achieves consistent and robust gains. The source code will be available at https://github.com/mona4399/ACR.