Out of distribution

However, using GANs to detect out-of-distribution instances by measuring the likelihood under the data distribution can fail (Nalisnick et al.,2019), while VAEs often generate ambiguous and blurry explanations. More recently, some re-searchers have argued that using auxiliary generative models in counterfactual generation incurs an engineering ... While out-of-distribution (OOD) generalization, robustness, and detection have been discussed in works related to reducing existential risks from AI (e.g., [Amodei et al., 2016, Hendrycks et al., 2022b]) the truth is that the vast majority of distribution shifts are not directly related to existential risks. cause of model crash under distribution shifts, they propose to realize out-of-distribution generalization by decorrelat-ing the relevant and irrelevant features. Since there is no extra supervision for separating relevant features from ir-relevant features, a conservative solution is to decorrelate all features. Feb 16, 2022 · To solve this critical problem, out-of-distribution (OOD) generalization on graphs, which goes beyond the I.I.D. hypothesis, has made great progress and attracted ever-increasing attention from the research community. In this paper, we comprehensively survey OOD generalization on graphs and present a detailed review of recent advances in this area. Sep 3, 2023 · Abstract. We study the out-of-distribution generalization of active learning that adaptively selects samples for annotation in learning the decision boundary of classification. Our empirical study finds that increasingly annotating seen samples may hardly benefit the generalization. To address the problem, we propose Counterfactual Active ... May 15, 2022 · 1. We propose an unsupervised method to distinguish in-distribution from out-of-distribution input. The results indicate that the assumptions and methods of outlier and deep anomaly detection are also relevant to the field of out-of-distribution detection. 2. The method works on the basis of an Isolation Forest. Oct 28, 2022 · Out-of-Distribution (OOD) detection separates ID (In-Distribution) data and OOD data from input data through a model. This problem has attracted increasing attention in the area of machine learning. OOD detection has achieved good intrusion detection, fraud detection, system health monitoring, sensor network event detection, and ecosystem interference detection. The method based on deep ... Hendrycks & Gimpel proposed a baseline method to detect out-of-distribution examples without further re-training networks. The method is based on an observation that a well-trained neural network tends to assign higher softmax scores to in-distribution examples than out-of-distribution Work done while at Cornell University. 1 marginal distribution of P X,Y for the input variable Xby P 0.Given a test input x ∈X, the problem of out-of-distribution detection can be formulated as a single-sample hypothesis testing task: H 0: x ∼P 0, vs. H 1: x ≁P 0. (1) Here the null hypothesis H 0 implies that the test input x is an in-distribution sample. The goal of trained in the closed-world setting, the out-of-distribution (OOD) issue arises and deteriorates customer experience when the models are deployed in production, facing inputs comingfromtheopenworld[9]. Forinstance,amodelmay wrongly but confidently classify an image of crab into the clappingclass,eventhoughnocrab-relatedconceptsappear in the ... Sep 15, 2022 · The unique contribution of this paper is two-fold, justified by extensive experiments. First, we present a realistic problem setting of OOD task for skin lesion. Second, we propose an approach to target the long-tailed and fine-grained aspects of the problem setting simultaneously to increase the OOD performance. Jan 22, 2019 · Out-of-distribution detection using an ensemble of self supervised leave-out classifiers A. Vyas, N. Jammalamadaka, X. Zhu, D. Das, B. Kaul, and T. L. Willke, “Out-of-distribution detection using an ensemble of self supervised leave-out classifiers,” in European Conference on Computer Vision, 2018, pp. 560–574. Dec 17, 2019 · The likelihood is dominated by the “background” pixels, whereas the likelihood ratio focuses on the “semantic” pixels and is thus better for OOD detection. Our likelihood ratio method corrects the background effect and significantly improves the OOD detection of MNIST images from an AUROC score of 0.089 to 0.994, based on a PixelCNN++ ... Dec 17, 2020 · While deep learning demonstrates its strong ability to handle independent and identically distributed (IID) data, it often suffers from out-of-distribution (OoD) generalization, where the test data come from another distribution (w.r.t. the training one). Designing a general OoD generalization framework to a wide range of applications is challenging, mainly due to possible correlation shift ... While out-of-distribution (OOD) generalization, robustness, and detection have been discussed in works related to reducing existential risks from AI (e.g., [Amodei et al., 2016, Hendrycks et al., 2022b]) the truth is that the vast majority of distribution shifts are not directly related to existential risks. 1ODIN: Out-of-DIstribution detector for Neural networks [21] failures are therefore often silent in that they do not result in explicit errors in the model. The above issue had been formulated as a problem of detecting whether an input data is from in-distribution (i.e. the training distribution) or out-of-distribution (i.e. a distri- Apr 19, 2023 · Recently, a class of compact and brain-inspired continuous-time recurrent neural networks has shown great promise in modeling autonomous navigation of ground ( 18, 19) and simulated drone vehicles end to end in a closed loop with their environments ( 21 ). These networks are called liquid time-constant (LTC) networks ( 35 ), or liquid networks. Jun 20, 2019 · To train our out-of-distribution detector, video features for unseen action categories are synthesized using generative adversarial networks trained on seen action category features. To the best of our knowledge, we are the first to propose an out-of-distribution detector based GZSL framework for action recognition in videos. Feb 1, 2023 · TL;DR: We propose a novel out-of-distribution detection method motivated by Modern Hopfield Energy, and futhur derive a simplified version that is effective, efficient and hyperparameter-free. Abstract : Out-of-Distribution (OOD) detection is essential for safety-critical applications of deep neural networks. cannot deliver reliable reasoning results when facing out-of-distribution samples. Next, even if supervision signals can be properly propagated between the neural and symbolic models, it is still possible that the NN predicts spurious fea-tures, leading to bad generalization performance (an exam-ple is provided in Sec. 6). Feb 21, 2022 · Most existing datasets with category and viewpoint labels 13,26,27,28 present two major challenges: (1) lack of control over the distribution of categories and viewpoints, or (2) small size. Thus ... CVF Open Access Feb 19, 2023 · Abstract. Recently, out-of-distribution (OOD) generalization has attracted attention to the robustness and generalization ability of deep learning based models, and accordingly, many strategies have been made to address different aspects related to this issue. However, most existing algorithms for OOD generalization are complicated and ... out-of-distribution examples, assuming our training set only contains older defendants referred as in-dis-tribution examples. The fractions of data are only for illustrative purposes. See details of in-distribution vs. out-of-distribution setup in §3.2. assistance, human-AI teams should outperform AI alone and human alone (e.g., in accuracy; also out-of-distribution examples, assuming our training set only contains older defendants referred as in-dis-tribution examples. The fractions of data are only for illustrative purposes. See details of in-distribution vs. out-of-distribution setup in §3.2. assistance, human-AI teams should outperform AI alone and human alone (e.g., in accuracy; also May 15, 2022 · 1. We propose an unsupervised method to distinguish in-distribution from out-of-distribution input. The results indicate that the assumptions and methods of outlier and deep anomaly detection are also relevant to the field of out-of-distribution detection. 2. The method works on the basis of an Isolation Forest. [ICML2022] Breaking Down Out-of-Distribution Detection: Many Methods Based on OOD Training Data Estimate a Combination of the Same Core Quantities [ICML2022] Scaling Out-of-Distribution Detection for Real-World Settings [ICML2022] POEM: Out-of-Distribution Detection with Posterior Sampling [NeurIPS2022] Deep Ensembles Work, But Are They Necessary? Mar 21, 2022 · Most of the existing Out-Of-Distribution (OOD) detection algorithms depend on single input source: the feature, the logit, or the softmax probability. However, the immense diversity of the OOD examples makes such methods fragile. There are OOD samples that are easy to identify in the feature space while hard to distinguish in the logit space and vice versa. Motivated by this observation, we ... Feb 19, 2023 · Abstract. Recently, out-of-distribution (OOD) generalization has attracted attention to the robustness and generalization ability of deep learning based models, and accordingly, many strategies have been made to address different aspects related to this issue. However, most existing algorithms for OOD generalization are complicated and ... Hendrycks & Gimpel proposed a baseline method to detect out-of-distribution examples without further re-training networks. The method is based on an observation that a well-trained neural network tends to assign higher softmax scores to in-distribution examples than out-of-distribution Work done while at Cornell University. 1 [ICML2022] Breaking Down Out-of-Distribution Detection: Many Methods Based on OOD Training Data Estimate a Combination of the Same Core Quantities [ICML2022] Scaling Out-of-Distribution Detection for Real-World Settings [ICML2022] POEM: Out-of-Distribution Detection with Posterior Sampling [NeurIPS2022] Deep Ensembles Work, But Are They Necessary? examples of 2 in-distribution (from CIFAR-100) and 1 out-of-distribution class (from CIFAR-10). The color coding shows the Mahalanobis outlier score, while the points are projections of embeddings of members of the in-distribution CIFAR-100 classes "sunflowers" (black plus signs) and "turtle" Jun 1, 2022 · In part I, we considered the case where we have a clean set of unlabelled data and must determine if a new sample comes from the same set. In part II, we considered the open-set recognition scenario where we also have class labels. This is particularly relevant to the real-world deployment of classifiers, which will inevitably encounter OOD data. A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks. Detecting test samples drawn sufficiently far away from the training distribution statistically or adversarially is a fundamental requirement for deploying a good classifier in many real-world machine learning applications. Oct 21, 2021 · Abstract: Out-of-distribution (OOD) detection is critical to ensuring the reliability and safety of machine learning systems. For instance, in autonomous driving, we would like the driving system to issue an alert and hand over the control to humans when it detects unusual scenes or objects that it has never seen during training time and cannot ... Out-of-Distribution (OOD) Detection with Deep Neural Networks based on PyTorch. and is designed such that it should be compatible with frameworks like pytorch-lightning and pytorch-segmentation-models . The library also covers some methods from closely related fields such as Open-Set Recognition, Novelty Detection, Confidence Estimation and ... Nov 11, 2021 · We propose Velodrome, a semi-supervised method of out-of-distribution generalization that takes labelled and unlabelled data from different resources as input and makes generalizable predictions. Out-of-distribution Neural networks and out-of-distribution data. A crucial criterion for deploying a strong classifier in many real-world... Out-of-Distribution (ODD). For Language and Vision activities, the term “distribution” has slightly different meanings. Various ODD detection techniques. This ... Sep 15, 2022 · The unique contribution of this paper is two-fold, justified by extensive experiments. First, we present a realistic problem setting of OOD task for skin lesion. Second, we propose an approach to target the long-tailed and fine-grained aspects of the problem setting simultaneously to increase the OOD performance. The outputs of an ensemble of networks can be used to estimate the uncertainty of a classifier. At test time, the estimated uncertainty for out-of-distribution samples turns out to be higher than the one for in-distribution samples. 3. level 2. AnvaMiba. cause of model crash under distribution shifts, they propose to realize out-of-distribution generalization by decorrelat-ing the relevant and irrelevant features. Since there is no extra supervision for separating relevant features from ir-relevant features, a conservative solution is to decorrelate all features. Feb 21, 2022 · It is well known that fine-tuning leads to better accuracy in-distribution (ID). However, in this paper, we find that fine-tuning can achieve worse accuracy than linear probing out-of-distribution (OOD) when the pretrained features are good and the distribution shift is large. On 10 distribution shift datasets (Breeds-Living17, Breeds-Entity30 ... The outputs of an ensemble of networks can be used to estimate the uncertainty of a classifier. At test time, the estimated uncertainty for out-of-distribution samples turns out to be higher than the one for in-distribution samples. 3. level 2. AnvaMiba. Feb 21, 2022 · Most existing datasets with category and viewpoint labels 13,26,27,28 present two major challenges: (1) lack of control over the distribution of categories and viewpoints, or (2) small size. Thus ... Jan 25, 2021 · The term 'out-of-distribution' (OOD) data refers to data that was collected at a different time, and possibly under different conditions or in a different environment, then the data collected to create the model. They may say that this data is from a 'different distribution'. Data that is in-distribution can be called novelty data. It is well known that fine-tuning leads to better accuracy in-distribution (ID). However, in this paper, we find that fine-tuning can achieve worse accuracy than linear probing out-of-distribution (OOD) when the pretrained features are good and the distribution shift is large. On 10 distribution shift datasets [ICML2022] Breaking Down Out-of-Distribution Detection: Many Methods Based on OOD Training Data Estimate a Combination of the Same Core Quantities [ICML2022] Scaling Out-of-Distribution Detection for Real-World Settings [ICML2022] POEM: Out-of-Distribution Detection with Posterior Sampling [NeurIPS2022] Deep Ensembles Work, But Are They Necessary? Dec 17, 2019 · The likelihood is dominated by the “background” pixels, whereas the likelihood ratio focuses on the “semantic” pixels and is thus better for OOD detection. Our likelihood ratio method corrects the background effect and significantly improves the OOD detection of MNIST images from an AUROC score of 0.089 to 0.994, based on a PixelCNN++ ... Feb 16, 2022 · To solve this critical problem, out-of-distribution (OOD) generalization on graphs, which goes beyond the I.I.D. hypothesis, has made great progress and attracted ever-increasing attention from the research community. In this paper, we comprehensively survey OOD generalization on graphs and present a detailed review of recent advances in this area. Jun 20, 2019 · To train our out-of-distribution detector, video features for unseen action categories are synthesized using generative adversarial networks trained on seen action category features. To the best of our knowledge, we are the first to propose an out-of-distribution detector based GZSL framework for action recognition in videos. out-of-distribution examples, assuming our training set only contains older defendants referred as in-dis-tribution examples. The fractions of data are only for illustrative purposes. See details of in-distribution vs. out-of-distribution setup in §3.2. assistance, human-AI teams should outperform AI alone and human alone (e.g., in accuracy; also A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks. Detecting test samples drawn sufficiently far away from the training distribution statistically or adversarially is a fundamental requirement for deploying a good classifier in many real-world machine learning applications. Mar 2, 2020 · Out-of-Distribution Generalization via Risk Extrapolation (REx) Distributional shift is one of the major obstacles when transferring machine learning prediction systems from the lab to the real world. To tackle this problem, we assume that variation across training domains is representative of the variation we might encounter at test time, but ... trained in the closed-world setting, the out-of-distribution (OOD) issue arises and deteriorates customer experience when the models are deployed in production, facing inputs comingfromtheopenworld[9]. Forinstance,amodelmay wrongly but confidently classify an image of crab into the clappingclass,eventhoughnocrab-relatedconceptsappear in the ... The outputs of an ensemble of networks can be used to estimate the uncertainty of a classifier. At test time, the estimated uncertainty for out-of-distribution samples turns out to be higher than the one for in-distribution samples. 3. level 2. AnvaMiba. ODIN: Out-of-Distribution Detector for Neural Networks Nov 11, 2021 · We propose Velodrome, a semi-supervised method of out-of-distribution generalization that takes labelled and unlabelled data from different resources as input and makes generalizable predictions. Let Dout denote an out-of-distribution dataset of (xout;y out)pairs where yout 2Y := fK+1;:::;K+Og;Yout\Yin =;. Depending on how different Dout is from Din, we categorize the OOD detection tasks into near-OOD and far-OOD. We first study the scenario where the model is fine-tuned only on the training set D in train without any access to OOD ... novelty detection (ND), open set recognition (OSR), out-of-distribution (OOD) detection, and outlier detection (OD). These sub-topics can be similar in the sense that they all define a certain in-distribution, with the common goal of detecting out-of-distribution samples under the open-world assumption. However, subtle differences exist among ... Aug 29, 2023 · ODIN is a preprocessing method for inputs that aims to increase the discriminability of the softmax outputs for In- and Out-of-Distribution data. Implements the Mahalanobis Method. Implements the Energy Score of Energy-based Out-of-distribution Detection. Uses entropy to detect OOD inputs. Implements the MaxLogit method. Hendrycks & Gimpel proposed a baseline method to detect out-of-distribution examples without further re-training networks. The method is based on an observation that a well-trained neural network tends to assign higher softmax scores to in-distribution examples than out-of-distribution Work done while at Cornell University. 1 [ICML2022] Breaking Down Out-of-Distribution Detection: Many Methods Based on OOD Training Data Estimate a Combination of the Same Core Quantities [ICML2022] Scaling Out-of-Distribution Detection for Real-World Settings [ICML2022] POEM: Out-of-Distribution Detection with Posterior Sampling [NeurIPS2022] Deep Ensembles Work, But Are They Necessary? out-of-distribution examples, assuming our training set only contains older defendants referred as in-dis-tribution examples. The fractions of data are only for illustrative purposes. See details of in-distribution vs. out-of-distribution setup in §3.2. assistance, human-AI teams should outperform AI alone and human alone (e.g., in accuracy; also To clarify the distinction between in-stock distribution, out-of-stock (OOS) distribution, and loss of distribution, it is essential to understand the dynamics of product availability and stock levels. Let’s refer to Exhibit 29.14, which provides an example of a brand’s incidence of purchase and stocks across four time periods. To clarify the distinction between in-stock distribution, out-of-stock (OOS) distribution, and loss of distribution, it is essential to understand the dynamics of product availability and stock levels. Let’s refer to Exhibit 29.14, which provides an example of a brand’s incidence of purchase and stocks across four time periods. this to be out-of-distribution clustering. Once a model Mhas been trained on the class homogeneity task, we can evaluate it for both out-of-distribution classification and out-of-distribution clustering. For the former, in which we are given x~ from a sample-label pair (~x;~y j~y = 2Y train), we can classify x~ by comparing it with samples of Jun 6, 2021 · Near out-of-distribution detection (OOD) is a major challenge for deep neural networks. We demonstrate that large-scale pre-trained transformers can significantly improve the state-of-the-art (SOTA) on a range of near OOD tasks across different data modalities. For instance, on CIFAR-100 vs CIFAR-10 OOD detection, we improve the AUROC from 85% (current SOTA) to more than 96% using Vision ... Nov 26, 2021 · Unsupervised out-of-distribution (U-OOD) detection has recently attracted much attention due its importance in mission-critical systems and broader applicability over its supervised counterpart. Despite this increase in attention, U-OOD methods suffer from important shortcomings. By performing a large-scale evaluation on different benchmarks and image modalities, we show in this work that most ... CVF Open Access Mar 21, 2022 · Most of the existing Out-Of-Distribution (OOD) detection algorithms depend on single input source: the feature, the logit, or the softmax probability. However, the immense diversity of the OOD examples makes such methods fragile. There are OOD samples that are easy to identify in the feature space while hard to distinguish in the logit space and vice versa. Motivated by this observation, we ... Out-of-distribution Neural networks and out-of-distribution data. A crucial criterion for deploying a strong classifier in many real-world... Out-of-Distribution (ODD). For Language and Vision activities, the term “distribution” has slightly different meanings. Various ODD detection techniques. This ... Jun 20, 2019 · To train our out-of-distribution detector, video features for unseen action categories are synthesized using generative adversarial networks trained on seen action category features. To the best of our knowledge, we are the first to propose an out-of-distribution detector based GZSL framework for action recognition in videos. Nov 11, 2021 · We propose Velodrome, a semi-supervised method of out-of-distribution generalization that takes labelled and unlabelled data from different resources as input and makes generalizable predictions. Aug 24, 2022 · We include results for four types of out-of-distribution samples: (1) dataset shift, where we evaluate the model on two other datasets with differences in the acquisition and population patterns (2) transformation shift where we apply artificial transformations to our ID data, (3) diagnostic shift, where we compare Covid-19 to non-Covid ... trained in the closed-world setting, the out-of-distribution (OOD) issue arises and deteriorates customer experience when the models are deployed in production, facing inputs comingfromtheopenworld[9]. Forinstance,amodelmay wrongly but confidently classify an image of crab into the clappingclass,eventhoughnocrab-relatedconceptsappear in the ... Jun 1, 2022 · In part I, we considered the case where we have a clean set of unlabelled data and must determine if a new sample comes from the same set. In part II, we considered the open-set recognition scenario where we also have class labels. This is particularly relevant to the real-world deployment of classifiers, which will inevitably encounter OOD data. out-of-distribution examples, assuming our training set only contains older defendants referred as in-dis-tribution examples. The fractions of data are only for illustrative purposes. See details of in-distribution vs. out-of-distribution setup in §3.2. assistance, human-AI teams should outperform AI alone and human alone (e.g., in accuracy; also out-of-distribution examples, assuming our training set only contains older defendants referred as in-dis-tribution examples. The fractions of data are only for illustrative purposes. See details of in-distribution vs. out-of-distribution setup in §3.2. assistance, human-AI teams should outperform AI alone and human alone (e.g., in accuracy; also Apr 19, 2023 · Recently, a class of compact and brain-inspired continuous-time recurrent neural networks has shown great promise in modeling autonomous navigation of ground ( 18, 19) and simulated drone vehicles end to end in a closed loop with their environments ( 21 ). These networks are called liquid time-constant (LTC) networks ( 35 ), or liquid networks. examples of 2 in-distribution (from CIFAR-100) and 1 out-of-distribution class (from CIFAR-10). The color coding shows the Mahalanobis outlier score, while the points are projections of embeddings of members of the in-distribution CIFAR-100 classes "sunflowers" (black plus signs) and "turtle" A project to improve out-of-distribution detection (open set recognition) and uncertainty estimation by changing a few lines of code in your project! Perform efficient inferences (i.e., do not increase inference time) without repetitive model training, hyperparameter tuning, or collecting additional data. machine-learning deep-learning pytorch ... We have summarized the main branches of works for Out-of-Distribution(OOD) Generalization problem, which are classified according to the research focus, including unsupervised representation learning, supervised learning models and optimization methods. For more details, please refer to our survey on OOD generalization. Out-of-distribution Neural networks and out-of-distribution data. A crucial criterion for deploying a strong classifier in many real-world... Out-of-Distribution (ODD). For Language and Vision activities, the term “distribution” has slightly different meanings. Various ODD detection techniques. This ... It is well known that fine-tuning leads to better accuracy in-distribution (ID). However, in this paper, we find that fine-tuning can achieve worse accuracy than linear probing out-of-distribution (OOD) when the pretrained features are good and the distribution shift is large. On 10 distribution shift datasets Towards Out-Of-Distribution Generalization: A Survey Jiashuo Liu*, Zheyan Shen∗, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, Peng Cui† Department of Computer Science and Technology Tsinghua University [email protected], [email protected], [email protected] Abstract ... trained in the closed-world setting, the out-of-distribution (OOD) issue arises and deteriorates customer experience when the models are deployed in production, facing inputs comingfromtheopenworld[9]. Forinstance,amodelmay wrongly but confidently classify an image of crab into the clappingclass,eventhoughnocrab-relatedconceptsappear in the ... [ICML2022] Breaking Down Out-of-Distribution Detection: Many Methods Based on OOD Training Data Estimate a Combination of the Same Core Quantities [ICML2022] Scaling Out-of-Distribution Detection for Real-World Settings [ICML2022] POEM: Out-of-Distribution Detection with Posterior Sampling [NeurIPS2022] Deep Ensembles Work, But Are They Necessary? out-of-distribution examples, assuming our training set only contains older defendants referred as in-dis-tribution examples. The fractions of data are only for illustrative purposes. See details of in-distribution vs. out-of-distribution setup in §3.2. assistance, human-AI teams should outperform AI alone and human alone (e.g., in accuracy; also A project to improve out-of-distribution detection (open set recognition) and uncertainty estimation by changing a few lines of code in your project! Perform efficient inferences (i.e., do not increase inference time) without repetitive model training, hyperparameter tuning, or collecting additional data. machine-learning deep-learning pytorch ... Nov 11, 2021 · We propose Velodrome, a semi-supervised method of out-of-distribution generalization that takes labelled and unlabelled data from different resources as input and makes generalizable predictions. CVF Open Access We have summarized the main branches of works for Out-of-Distribution(OOD) Generalization problem, which are classified according to the research focus, including unsupervised representation learning, supervised learning models and optimization methods. For more details, please refer to our survey on OOD generalization. Jun 20, 2019 · To train our out-of-distribution detector, video features for unseen action categories are synthesized using generative adversarial networks trained on seen action category features. To the best of our knowledge, we are the first to propose an out-of-distribution detector based GZSL framework for action recognition in videos. Mar 25, 2022 · All solutions mentioned above, such as regularization, multimodality, scaling, and invariant risk minimization, can improve distribution shift and out-of-distribution generalization, ultimately ... Towards Out-Of-Distribution Generalization: A Survey Jiashuo Liu*, Zheyan Shen∗, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, Peng Cui† Department of Computer Science and Technology Tsinghua University [email protected], [email protected], [email protected] Abstract ... Feb 21, 2022 · It is well known that fine-tuning leads to better accuracy in-distribution (ID). However, in this paper, we find that fine-tuning can achieve worse accuracy than linear probing out-of-distribution (OOD) when the pretrained features are good and the distribution shift is large. On 10 distribution shift datasets (Breeds-Living17, Breeds-Entity30 ... We evaluate our method on a diverse set of in- and out-of-distribution dataset pairs. In many settings, our method outperforms other methods by a large margin. The contri-butions of our paper are summarized as follows: • We propose a novel experimental setting and a novel training methodology for out-of-distribution detection in neural networks. examples of 2 in-distribution (from CIFAR-100) and 1 out-of-distribution class (from CIFAR-10). The color coding shows the Mahalanobis outlier score, while the points are projections of embeddings of members of the in-distribution CIFAR-100 classes "sunflowers" (black plus signs) and "turtle" In-distribution Out-of-distribution Figure 1. Learned confidence estimates can be used to easily sep-arate in- and out-of-distribution examples. Here, the CIFAR-10 test set is used as the in-distribution dataset, and TinyImageNet, LSUN, and iSUN are used as the out-of-distribution datasets. The model is trained using a DenseNet architecture. | Cofnkqrzsnwt (article) | Msvinjk.