OOD Detection Research Articles

Out-of-Distribution Detection with Memory-Augmented Variational Autoencoder

This paper proposes a novel method capable of both detecting OOD data and generating in-distribution data samples. To achieve this, a VAE model is adopted and augmented with a memory module, providing capacities for identifying OOD data and synthesising new in-distribution samples. The proposed VAE is trained on normal data and the memory stores prototypical patterns of the normal data distribution. At test time, the input is encoded by the VAE encoder; this encoding is used as a query to retrieve related memory items, which are then integrated with the input encoding and passed to the decoder for reconstruction. Normal samples reconstruct well and yield low reconstruction errors, while OOD inputs produce high reconstruction errors as their encodings get replaced by retrieved normal patterns. Prior works use memory modules for OOD detection with autoencoders, but this method leverages a VAE architecture to enable generation abilities. Experiments conducted with CIFAR-10 and MNIST datasets show that the memory-augmented VAE consistently outperforms the baseline, particularly where OOD data resembles normal patterns. This notable improvement is due to the enhanced latent space representation provided by the VAE. Overall, the memory-equipped VAE framework excels in identifying OOD and generating creative examples effectively.

Resource‐adaptive and OOD‐robust inference of deep neural networks on IoT devices

AbstractEfficiently executing inference tasks of deep neural networks on devices with limited resources poses a significant load in IoT systems. To alleviate the load, one innovative method is branching that adds extra layers with classification exits to a pre‐trained model, enabling inputs with high‐confidence predictions to exit early, thus reducing inference cost. However, branching networks, not originally tailored for IoT environments, are susceptible to noisy and out‐of‐distribution (OOD) data, and they demand additional training for optimal performance. The authors introduce BrevisNet, a novel branching methodology designed for creating on‐device branching models that are both resource‐adaptive and noise‐robust for IoT applications. The method leverages the refined uncertainty estimation capabilities of Dirichlet distributions for classification predictions, combined with the superior OOD detection of energy‐based models. The authors propose a unique training approach and thresholding technique that enhances the precision of branch predictions, offering robustness against noise and OOD inputs. The findings demonstrate that BrevisNet surpasses existing branching techniques in training efficiency, accuracy, overall performance, and robustness.

Can input reconstruction be used to directly estimate uncertainty of a dose prediction U-Net model?

The reliable and efficient estimation of uncertainty in artificial intelligence (AI) models poses an ongoing challenge in many fields such as radiation therapy. AI models are intended to automate manual steps involved in the treatment planning workflow. We focus in this study on dose prediction models that predict an optimal dose trade-off for each new patient for a specific treatment modality. They can guide physicians in the optimization, be part of automatic treatment plan generation or support decision in treatmentindication. Most common uncertainty estimation methods are based on Bayesian approximations, like Monte Carlo dropout (MCDO) or Deep ensembling (DE). These two techniques, however, have a high inference time (i.e., require multiple inference passes) and might not work for detecting out-of-distribution (OOD) data (i.e., overlapping uncertainty estimate for in-distribution (ID) and OOD). In this study, we present a direct uncertainty estimation method and apply it for a dose prediction U-Net architecture. It can be used to flag OOD data and give information on the quality of the doseprediction. Our method consists in the addition of a branch decoding from the bottleneck which reconstructs the CT scan given as input. The input reconstruction error can be used as a surrogate of the model uncertainty. For the proof-of-concept, our method is applied to proton therapy dose prediction in head and neck cancer patients. A dataset of 60 oropharyngeal patients was used to train the network using a nested cross-validation approach with 11 folds (training: 50 patients, validation: 5 patients, test: 5 patients). For the OOD experiment, we used 10 extra patients with a different head and neck sub-location. Accuracy, time-gain, and OOD detection are analyzed for our method in this particular application and compared with the popular MCDO andDE. The additional branch did not reduce the accuracy of the dose prediction model. The median absolute error is close to zero for the target volumes and less than 1% of the dose prescription for organs at risk. Our input reconstruction method showed a higher Pearson correlation coefficient with the prediction error (0.620) than DE (0.447) and MCDO (between 0.599 and 0.612). Moreover, our method allows an easier identification of OOD (no overlap for ID and OOD data and a Z-score of 34.05). The uncertainty is estimated simultaneously to the regression task, therefore requires less time and computationalresources. This study shows that the error in the CT scan reconstruction can be used as a surrogate of the uncertainty of the model. The Pearson correlation coefficient with the dose prediction error is slightly higher than state-of-the-art techniques. OOD data can be more easily detected and the uncertainty metric is computed simultaneously to the regression task, therefore faster than MCDO or DE. The code and pretrained model are available on the gitlab repository: https://gitlab.com/ai4miro/ct-reconstruction-for-uncertainty-quatification-of-hdunet.

Meta transfer evidence deep learning for trustworthy few-shot classification

Although the standard meta-learning methods have demonstrated strong performance in few-shot image classification scenarios, the models typically lack the capability to assess the reliability of their predictions, which can lead to risks in certain applications. Aiming at this problem, we first propose a meta-learning-based Evidential Deep Learning (EDL) called Meta Evidence Deep Learning (MetaEDL), which enables reliable prediction in the few-show image classification scenario. Being the same as general meta-learning methods, MetaEDL commonly employed shallow neural networks as feature extractors to avoid overfitting when dealing with few-shot samples, which significantly restricts the model’s ability to extract features. To further address this limitation, we propose a Meta Transfer Evidence Deep Learning (MetaTEDL) to address the few-shot trustworthy classification issue. MetaTEDL adopts a large-scale pre-trained neural network as its feature extractor. In the meta-training process, we only train two lightweight neuron operations Scaling and Shifting to reduce the risk of over-fitting. Then, two evidential head neural networks are trained to integrate evidence from different sources, aiming to improve the quality of the evidence output. We conduct comprehensive experiments on several challenging few-shot classification benchmarks. The results indicate that our proposed method not only outperforms other conventional meta-learning methods in terms of few-shot classification performance, but also has good UQ (uncertainty quantification), Uncertainty-guided active learning, and OOD (Out-of-Distribution) detection capabilities.

Out-of-Distribution Detection with Virtual Outlier Smoothing

AbstractDetecting out-of-distribution (OOD) inputs plays a crucial role in guaranteeing the reliability of deep neural networks (DNNs) when deployed in real-world scenarios. However, DNNs typically exhibit overconfidence in OOD samples, which is attributed to the similarity in patterns between OOD and in-distribution (ID) samples. To mitigate this overconfidence, advanced approaches suggest the incorporation of auxiliary OOD samples during model training, where the outliers are assigned with an equal likelihood of belonging to any category. However, identifying outliers that share patterns with ID samples poses a significant challenge. To address the challenge, we propose a novel method, Virtual Outlier Smoothing (VOSo), which constructs auxiliary outliers using ID samples, thereby eliminating the need to search for OOD samples. Specifically, VOSo creates these virtual outliers by perturbing the semantic regions of ID samples and infusing patterns from other ID samples. For instance, a virtual outlier might consist of a cat’s face with a dog’s nose, where the cat’s face serves as the semantic feature for model prediction. Meanwhile, VOSo adjusts the labels of virtual OOD samples based on the extent of semantic region perturbation, aligning with the notion that virtual outliers may contain ID patterns. Extensive experiments are conducted on diverse OOD detection benchmarks, demonstrating the effectiveness of the proposed VOSo. Our code will be available at https://github.com/junz-debug/VOSo.

Towards trustworthy dataset distillation

Efficiency and trustworthiness are two eternal pursuits when applying deep learning in practical scenarios. Considering efficiency, dataset distillation (DD) endeavors to reduce training costs by distilling large datasets into tiny ones. However, existing methods concentrate on in-distribution (InD) classification, disregarding out-of-distribution (OOD) samples. On the other hand, OOD detection aims to enhance models’ trustworthiness, which is inefficiently achieved in full-data settings. For the first time, we consider both issues and propose a novel paradigm called Trustworthy Dataset Distillation (TrustDD). By distilling both InD samples and outliers, the condensed datasets are capable of training models competent in both InD classification and OOD detection. To alleviate the requirement of real outlier data, we further propose to corrupt InD samples to generate pseudo-outliers for TrustDD, namely Pseudo-Outlier Exposure. Comprehensive experiments demonstrate the effectiveness of TrustDD. TrustDD is more trustworthy and applicable to real open-world scenarios. Our code is available at https://github.com/mashijie1028/TrustDD.

The impact of fine-tuning paradigms on unknown plant diseases recognition

Plant diseases pose significant threats to agriculture, impacting both food safety and public health. Traditional plant disease detection systems are typically limited to recognizing disease categories included in the training dataset, rendering them ineffective against new disease types. Although out-of-distribution (OOD) detection methods have been proposed to address this issue, the impact of fine-tuning paradigms on these methods has been overlooked. This paper focuses on studying the impact of fine-tuning paradigms on the performance of detecting unknown plant diseases. Currently, fine-tuning on visual tasks is mainly divided into visual-based models and visual-language-based models. We first discuss the limitations of large-scale visual language models in this task: textual prompts are difficult to design. To avoid the side effects of textual prompts, we futher explore the effectiveness of purely visual pre-trained models for OOD detection in plant disease tasks. Specifically, we employed five publicly accessible datasets to establish benchmarks for open-set recognition, OOD detection, and few-shot learning in plant disease recognition. Additionally, we comprehensively compared various OOD detection methods, fine-tuning paradigms, and factors affecting OOD detection performance, such as sample quantity. The results show that visual prompt tuning outperforms fully fine-tuning and linear probe tuning in out-of-distribution detection performance, especially in the few-shot scenarios. Notably, the max-logit-based on visual prompt tuning achieves an AUROC score of 94.8%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document} in the 8-shot setting, which is nearly comparable to the method of fully fine-tuning on the full dataset (95.2%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document}), which implies that an appropriate fine-tuning paradigm can directly improve OOD detection performance. Finally, we visualized the prediction distributions of different OOD detection methods and discussed the selection of thresholds. Overall, this work lays the foundation for unknown plant disease recognition, providing strong support for the security and reliability of plant disease recognition systems. We will release our code at https://github.com/JiuqingDong/PDOOD to further advance this field.

Uncertainty-aware multiple-instance learning for reliable classification: Application to optical coherence tomography

Deep learning classification models for medical image analysis often perform well on data from scanners that were used to acquire the training data. However, when these models are applied to data from different vendors, their performance tends to drop substantially. Artifacts that only occur within scans from specific scanners are major causes of this poor generalizability. We aimed to enhance the reliability of deep learning classification models using a novel method called Uncertainty-Based Instance eXclusion (UBIX). UBIX is an inference-time module that can be employed in multiple-instance learning (MIL) settings. MIL is a paradigm in which instances (generally crops or slices) of a bag (generally an image) contribute towards a bag-level output. Instead of assuming equal contribution of all instances to the bag-level output, UBIX detects instances corrupted due to local artifacts on-the-fly using uncertainty estimation, reducing or fully ignoring their contributions before MIL pooling. In our experiments, instances are 2D slices and bags are volumetric images, but alternative definitions are also possible. Although UBIX is generally applicable to diverse classification tasks, we focused on the staging of age-related macular degeneration in optical coherence tomography. Our models were trained on data from a single scanner and tested on external datasets from different vendors, which included vendor-specific artifacts. UBIX showed reliable behavior, with a slight decrease in performance (a decrease of the quadratic weighted kappa (κw) from 0.861 to 0.708), when applied to images from different vendors containing artifacts; while a state-of-the-art 3D neural network without UBIX suffered from a significant detriment of performance (κw from 0.852 to 0.084) on the same test set. We showed that instances with unseen artifacts can be identified with OOD detection. UBIX can reduce their contribution to the bag-level predictions, improving reliability without retraining on new data. This potentially increases the applicability of artificial intelligence models to data from other scanners than the ones for which they were developed. The source code for UBIX, including trained model weights, is publicly available through https://github.com/qurAI-amsterdam/ubix-for-reliable-classification.

A holistic semi-supervised method for imbalanced fault diagnosis of rotational machinery with out-of-distribution samples

Fault diagnosis plays a critical role in ensuring the reliability and safety of industrial systems. Despite the success of semi-supervised learning in fault diagnosis, challenges persist in dealing with imbalanced datasets and out-of-distribution (OOD) samples, particularly in industrial settings where fault incidents are rare. The rarity of fault incidents in industrial contexts leads to dataset imbalances, while changes in conditions may introduce OOD samples. To address these challenges, this study proposes a novel semi-supervised fault diagnosis method that simultaneously considers the imbalanced distribution of fault categories and OOD samples. Firstly, the semi-supervised framework has been presented, which combines pseudo-labeling and consistency regularization techniques. Then, the self-adaptive threshold has been devised to balance confidence levels among fault classes using dataset-specific and class-specific thresholds, thereby enhancing the model's robustness in real-world scenarios. Additionally, the OOD detection strategy based on contrastive learning has been proposed, wherein predictions from two different views of each sample estimate its “likelihood” of belonging to in-distribution (ID) or OOD categories. This strategy improves system safety by alerting operators to potential anomalies. Comprehensive experimental validations and comparisons were conducted on two rotational machinery fault datasets, demonstrating that the proposed method achieves superior performance compared to the alternative methods and is applicable for practical application.

When an extra rejection class meets out-of-distribution detection in long-tailed image classification

Detecting Out-of-Distribution (OOD) inputs is essential for reliable deep learning in the open world. However, most existing OOD detection methods have been developed based on training sets that exhibit balanced class distributions, making them susceptible when confronted with training sets following a long-tailed distribution. To alleviate this problem, we propose an effective three-branch training framework, which demonstrates the efficacy of incorporating an extra rejection class along with auxiliary outlier training data for effective OOD detection in long-tailed image classification. In our proposed framework, all outlier training samples are assigned the label of the rejection class. We employ an inlier loss, an outlier loss, and a Tail-class prototype induced Supervised Contrastive Loss (TSCL) to train both the in-distribution classifier and OOD detector within one network. During inference, the OOD detector is constructed using the rejection class. Extensive experimental results demonstrate that the superior OOD detection performance of our proposed method in long-tailed image classification. For example, in the more challenging case where CIFAR100-LT is used as in-distribution, our method improves the average AUROC by 1.23% and reduces the average FPR95 by 3.18% compared to the baseline method utilizing Outlier Exposure (OE). Code is available at github.

EEG 기반 감정 분류에서 MSP를 사용한 OOD 검출 적용

EEG 기반 감정 분류에서 MSP를 사용한 OOD 검출 적용

Leveraging diffusion models for unsupervised out-of-distribution detection on image manifold.

Out-of-distribution (OOD) detection is crucial for enhancing the reliability of machine learning models when confronted with data that differ from their training distribution. In the image domain, we hypothesize that images inhabit manifolds defined by latent properties such as color, position, and shape. Leveraging this intuition, we propose a novel approach to OOD detection using a diffusion model to discern images that deviate from the in-domain distribution. Our method involves training a diffusion model using in-domain images. At inference time, we lift an image from its original manifold using a masking process, and then apply a diffusion model to map it towards the in-domain manifold. We measure the distance between the original and mapped images, and identify those with a large distance as OOD. Our experiments encompass comprehensive evaluation across various datasets characterized by differences in color, semantics, and resolution. Our method demonstrates strong and consistent performance in detecting OOD images across the tested datasets, highlighting its effectiveness in handling images with diverse characteristics. Additionally, ablation studies confirm the significant contribution of each component in our framework to the overall performance.

Revisiting Confidence Estimation: Towards Reliable Failure Prediction.

Reliable confidence estimation is a challenging yet fundamental requirement in many risk-sensitive applications. However, modern deep neural networks are often overconfident for their incorrect predictions, i.e., misclassified samples from known classes, and out-of-distribution (OOD) samples from unknown classes. In recent years, many confidence calibration and OOD detection methods have been developed. In this paper, we find a general, widely existing but actually-neglected phenomenon that most confidence estimation methods are harmful for detecting misclassification errors. We investigate this problem and reveal that popular calibration and OOD detection methods often lead to worse confidence separation between correctly classified and misclassified examples, making it difficult to decide whether to trust a prediction or not. Finally, we propose to enlarge the confidence gap by finding flat minima, which yields state-of-the-art failure prediction performance under various settings including balanced, long-tailed, and covariate-shift classification scenarios. Our study not only provides a strong baseline for reliable confidence estimation but also acts as a bridge between understanding calibration, OOD detection, and failure prediction.

Augmenting the Softmax with Additional Confidence Scores for Improved Selective Classification with Out-of-Distribution Data

Detecting out-of-distribution (OOD) data is a task that is receiving an increasing amount of research attention in the domain of deep learning for computer vision. However, the performance of detection methods is generally evaluated on the task in isolation, rather than also considering potential downstream tasks in tandem. In this work, we examine selective classification in the presence of OOD data (SCOD). That is to say, the motivation for detecting OOD samples is to reject them so their impact on the quality of predictions is reduced. We show under this task specification, that existing post-hoc methods perform quite differently compared to when evaluated only on OOD detection. This is because it is no longer an issue to conflate in-distribution (ID) data with OOD data if the ID data is going to be misclassified. However, the conflation within ID data of correct and incorrect predictions becomes undesirable. We also propose a novel method for SCOD, Softmax Information Retaining Combination (SIRC), that augments a softmax-based confidence score with a secondary class-agnostic feature-based score. Thus, the ability to identify OOD samples is improved without sacrificing separation between correct and incorrect ID predictions. Experiments on a wide variety of ImageNet-scale datasets and convolutional neural network architectures show that SIRC is able to consistently match or outperform the baseline for SCOD, whilst existing OOD detection methods fail to do so. Interestingly, we find that the secondary scores investigated for SIRC do not consistently improve performance on all tested OOD datasets. To address this issue, we further extend SIRC to incorporate multiple secondary scores (SIRC+). This further improves SCOD performance, both generally, and in terms of consistency over diverse distribution shifts. Code is available at https://github.com/Guoxoug/SIRC.

Exploring Channel-Aware Typical Features for Out-of-Distribution Detection

Detecting out-of-distribution (OOD) data is essential to ensure the reliability of machine learning models when deployed in real-world scenarios. Different from most previous test-time OOD detection methods that focus on designing OOD scores, we delve into the challenges in OOD detection from the perspective of typicality and regard the feature’s high-probability region as the feature’s typical set. However, the existing typical-feature-based OOD detection method implies an assumption: the proportion of typical feature sets for each channel is fixed. According to our experimental analysis, each channel contributes differently to OOD detection. Adopting a fixed proportion for all channels results in several channels losing too many typical features or incorporating too many abnormal features, resulting in low performance. Therefore, exploring the channel-aware typical features is crucial to better-separating ID and OOD data. Driven by this insight, we propose expLoring channel-Aware tyPical featureS (LAPS). Firstly, LAPS obtains the channel-aware typical set by calibrating the channel-level typical set with the global typical set from the mean and standard deviation. Then, LAPS rectifies the features into channel-aware typical sets to obtain channel-aware typical features. Finally, LAPS leverages the channel-aware typical features to calculate the energy score for OOD detection. Theoretical and visual analyses verify that LAPS achieves a better bias-variance trade-off. Experiments verify the effectiveness and generalization of LAPS under different architectures and OOD scores.

CASE: Exploiting Intra-class Compactness and Inter-class Separability of Feature Embeddings for Out-of-Distribution Detection

Detecting out-of-distribution (OOD) inputs is critical for reliable machine learning, but deep neural networks often make overconfident predictions, even for OOD inputs that deviate from the distribution of training data. Prior methods relied on the widely used softmax cross-entropy (CE) loss that is adequate for classifying in-distribution (ID) samples but not optimally designed for OOD detection. To address this issue, we propose CASE, a simple and effective OOD detection method by explicitly improving intra-class Compactness And inter-class Separability of feature Embeddings. To enhance the separation between ID and OOD samples, CASE uses a dual-loss framework, which includes a separability loss that maximizes the inter-class Euclidean distance to promote separability among different class centers, along with a compactness loss that minimizes the intra-class Euclidean distance to encourage samples to be close to their class centers. In particular, the class centers are defined as a free optimization parameter of the model and updated by gradient descent, which is simple and further enhances the OOD detection performance. Extensive experiments demonstrate the superiority of CASE, which reduces the average FPR95 by 37.11% and improves the average AUROC by 15.89% compared to the baseline method using a softmax confidence score on the more challenging CIFAR-100 model.

Unsupervised Layer-Wise Score Aggregation for Textual OOD Detection

Out-of-distribution (OOD) detection is a rapidly growing field due to new robustness and security requirements driven by an increased number of AI-based systems. Existing OOD textual detectors often rely on anomaly scores (\textit{e.g.}, Mahalanobis distance) computed on the embedding output of the last layer of the encoder. In this work, we observe that OOD detection performance varies greatly depending on the task and layer output. More importantly, we show that the usual choice (the last layer) is rarely the best one for OOD detection and that far better results can be achieved, provided that an oracle selects the best layer. We propose a data-driven, unsupervised method to leverage this observation to combine layer-wise anomaly scores. In addition, we extend classical textual OOD benchmarks by including classification tasks with a more significant number of classes (up to 150), which reflects more realistic settings. On this augmented benchmark, we show that the proposed post-aggregation methods achieve robust and consistent results comparable to using the best layer according to an oracle while removing manual feature selection altogether.

TagFog: Textual Anchor Guidance and Fake Outlier Generation for Visual Out-of-Distribution Detection

Out-of-distribution (OOD) detection is crucial in many real-world applications. However, intelligent models are often trained solely on in-distribution (ID) data, leading to overconfidence when misclassifying OOD data as ID classes. In this study, we propose a new learning framework which leverage simple Jigsaw-based fake OOD data and rich semantic embeddings (`anchors') from the ChatGPT description of ID knowledge to help guide the training of the image encoder. The learning framework can be flexibly combined with existing post-hoc approaches to OOD detection, and extensive empirical evaluations on multiple OOD detection benchmarks demonstrate that rich textual representation of ID knowledge and fake OOD knowledge can well help train a visual encoder for OOD detection. With the learning framework, new state-of-the-art performance was achieved on all the benchmarks. The code is available at https://github.com/Cverchen/TagFog.

Learning by Erasing: Conditional Entropy Based Transferable Out-of-Distribution Detection

Detecting OOD inputs is crucial to deploy machine learning models to the real world safely. However, existing OOD detection methods require an in-distribution (ID) dataset to retrain the models. In this paper, we propose a Deep Generative Models (DGMs) based transferable OOD detection that does not require retraining on the new ID dataset. We first establish and substantiate two hypotheses on DGMs: DGMs exhibit a predisposition towards acquiring low-level features, in preference to semantic information; the lower bound of DGM's log-likelihoods is tied to the conditional entropy between the model input and target output. Drawing on the aforementioned hypotheses, we present an innovative image-erasing strategy, which is designed to create distinct conditional entropy distributions for each individual ID dataset. By training a DGM on a complex dataset with the proposed image-erasing strategy, the DGM could capture the discrepancy of conditional entropy distribution for varying ID datasets, without re-training. We validate the proposed method on the five datasets and show that, without retraining, our method achieves comparable performance to the state-of-the-art group-based OOD detection methods. The project codes will be open-sourced on our project website.

How to Overcome Curse-of-Dimensionality for Out-of-Distribution Detection?

Machine learning models deployed in the wild can be challenged by out-of-distribution (OOD) data from unknown classes. Recent advances in OOD detection rely on distance measures to distinguish samples that are relatively far away from the in-distribution (ID) data. Despite the promise, distance-based methods can suffer from the curse-of-dimensionality problem, which limits the efficacy in high dimensional feature space. To combat this problem, we propose a novel framework, Subspace Nearest Neighbor (SNN), for OOD detection. In training, our method regularizes the model and its feature representation by leveraging the most relevant subset of dimensions (i.e. subspace). The subspace learning yields highly distinguishable distance measures between ID and OOD data. We provide comprehensive experiments and ablations to validate the efficacy of SNN. Compared to the current best distance-based method, SNN reduces the average FPR95 by 15.96% on the CIFAR-100 benchmark.