Label-Assisted Memory Autoencoder for Unsupervised Out-of-Distribution Detection

Zhang, Shuyi; Pan, Chao; Song, Liying; Wu, Xiaoyu; Hu, Zheng; Pei, Ke; Tiňo, Peter; Yao, Xin

doi:10.1007/978-3-030-86523-8_48

Cited by 4 publications

(5 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Similarly, the overconfident phenomenon is also reported in (Choi, Jang, and Alemi 2018;Denouden et al 2018;Gong et al 2019;Nalisnick et al 2018;Zhang et al 2021). Lastly, the reconstruction error distribution of "hard" OODs which contain richer contents and diverse pixels compared to ID is skewed to the right side as expected, and this is consistent with the reconstruction error assumption.…”

Section: Adjustment Coefficientsupporting

confidence: 79%

“…(b) an autoencoder (AE), including encoder f enc to compress high-dimensional data features with parameters θ enc and deconder f dec recreating x denoted by x from the latent representation z with parameters θ dec . Different from works (Oza and Patel 2019; Zhang et al 2021) integrating classifier and autoencoder in one hybrid model simultaneously and utilizing raw pixels reconstruction error as score function, the CLF and AE modules in our method are independent of each other. Furthermore, we transform the reconstruction error to CLF latent space instead of pixels space for further aggregation.…”

Section: Overall Conceptmentioning

confidence: 97%

“…This problem is addressed by considering a likelihood ratio (Ren et al 2019), taking the input complexity into account (Serrà et al 2019), and likelihood regret (Xiao, Yan, and Amit 2020). For autoencoder, the specific OOD inputs can be reconstructed well demonstrated by (Denouden et al 2018;Gong et al 2019;Zhang et al 2021). In contrast, we transform the reconstruction error to latent space of classifier as a supplement and propose an adjustment coefficient to alleviate the overconfidence problem.…”

Section: Out-of-distribution Detection For Generative Modelsmentioning

confidence: 99%

“…Overall, the transformed reconstruction error exhibits competitive OOD detection performance compared with the raw pixels. Based on the same reconstruction error assumption, Gong et al (Gong et al 2019) and Zhang et al (Zhang et al 2021) incorporate memory module to autoencoder and directly use raw pixels reconstruction error to detect OOD samples. However, they find that this assumption does not always hold and the autoencoder can reconstruct specific OOD data well with low reconstruction error.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

READ: Aggregating Reconstruction Error into Out-of-Distribution Detection

Jiang

Cheng

et al. 2023

AAAI

View full text Add to dashboard Cite

Detecting out-of-distribution (OOD) samples is crucial to the safe deployment of a classifier in the real world. However, deep neural networks are known to be overconfident for abnormal data. Existing works directly design score function by mining the inconsistency from classifier for in-distribution (ID) and OOD. In this paper, we further complement this inconsistency with reconstruction error, based on the assumption that an autoencoder trained on ID data cannot reconstruct OOD as well as ID. We propose a novel method, READ (Reconstruction Error Aggregated Detector), to unify inconsistencies from classifier and autoencoder. Specifically, the reconstruction error of raw pixels is transformed to latent space of classifier. We show that the transformed reconstruction error bridges the semantic gap and inherits detection performance from the original. Moreover, we propose an adjustment strategy to alleviate the overconfidence problem of autoencoder according to a fine-grained characterization of OOD data. Under two scenarios of pre-training and retraining, we respectively present two variants of our method, namely READ-MD (Mahalanobis Distance) only based on pre-trained classifier and READ-ED (Euclidean Distance) which retrains the classifier. Our methods do not require access to test time OOD data for fine-tuning hyperparameters. Finally, we demonstrate the effectiveness of the proposed methods through extensive comparisons with state-of-the-art OOD detection algorithms. On a CIFAR-10 pre-trained WideResNet, our method reduces the average FPR@95TPR by up to 9.8% compared with previous state-of-the-art.

show abstract

Section: Adjustment Coefficientsupporting

confidence: 79%

Section: Overall Conceptmentioning

confidence: 97%

Section: Out-of-distribution Detection For Generative Modelsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

READ: Aggregating Reconstruction Error into Out-of-Distribution Detection

Jiang

Cheng

et al. 2023

AAAI

View full text Add to dashboard Cite

show abstract

“…Secondly, introducing additional loss term significantly deviates the training objective from its original task, i.e., to obtain satisfactory classification. To alleviate these constraints, few works in the literature have explored unsupervised detection of OOD samples, i.e., without using any OOD sample during training phase [7,8]. One straightforward approach is to use the softmax value as an indicator of OOD detection [9,10].…”

Section: Introductionmentioning

confidence: 99%

Compact Feature Representation for Unsupervised Ood Detection

Saha

Gawlikowski

Nandy

et al. 2022

IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium

View full text Add to dashboard Cite

Distributional mismatch between training and test data may cause the remote sensing models to behave in unpredictable manner, thus reducing the trustworthiness of such models. Most existing methods for out-of-distribution (OOD) detection rely on availability of OOD samples during training. However, access to OOD data during training is counter intuitive and may be impractical sometimes. Considering this, we propose an unsupervised OOD detection model that does not require training OOD data. The proposed method works by projecting the in-domain samples as a union of 1-dimensional subspaces. Due to the compact feature representation of indomain samples, OOD samples are less likely to occupy the same feature space, thus they are easily identified. Experimental results demonstrate the capability of the proposed method to detect OOD samples.

show abstract