Open Set Recognition by Regularising Classifier with Fake Data Generated by Generative Adversarial Networks

Jo, Inhyuk; Kim, Jungtaek; Kang, Hyohyeong; Kim, Yong-Deok; Choi, Seungjin

doi:10.1109/icassp.2018.8461700

Cited by 36 publications

(19 citation statements)

References 4 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Many techniques have been proposed for OOD detection based on traditional classification algorithms, such as OCSVM [ 41 , 43 , 44 , 45 ], KNN [ 15 ], and Kmeans [ 16 ]. Recently, deep learning-based OOD-detection techniques have been proposed [ 17 , 19 , 20 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 46 , 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 ], some of which we now highlight.…”

Section: Background and Related Workmentioning

confidence: 99%

“…GANs are generally incorporated into an open set discriminative framework by training a GAN on the known input dataset and converting the trained discriminator into an OOD classifier [ 46 , 59 ]. Alternatively, the generator used to synthesize samples may designate them as from an OOD class, and then add them to the full training set used to train a classifier [ 47 , 48 , 49 , 50 ]. However, such GAN augmentation techniques may produce samples that do not adequately cover the entirety of in-distribution (IIN) decision boundaries, resulting in non-optimal OOD classifiers [ 60 ].…”

Section: Background and Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Out-of-Distribution Detection of Human Activity Recognition with Smartwatch Inertial Sensors

Boyer

Burns

Whyne

2021

Sensors

View full text Add to dashboard Cite

Out-of-distribution (OOD) in the context of Human Activity Recognition (HAR) refers to data from activity classes that are not represented in the training data of a Machine Learning (ML) algorithm. OOD data are a challenge to classify accurately for most ML algorithms, especially deep learning models that are prone to overconfident predictions based on in-distribution (IIN) classes. To simulate the OOD problem in physiotherapy, our team collected a new dataset (SPARS9x) consisting of inertial data captured by smartwatches worn by 20 healthy subjects as they performed supervised physiotherapy exercises (IIN), followed by a minimum 3 h of data captured for each subject as they engaged in unrelated and unstructured activities (OOD). In this paper, we experiment with three traditional algorithms for OOD-detection using engineered statistical features, deep learning-generated features, and several popular deep learning approaches on SPARS9x and two other publicly-available human activity datasets (MHEALTH and SPARS). We demonstrate that, while deep learning algorithms perform better than simple traditional algorithms such as KNN with engineered features for in-distribution classification, traditional algorithms outperform deep learning approaches for OOD detection for these HAR time series datasets.

show abstract

Section: Background and Related Workmentioning

confidence: 99%

Section: Background and Related Workmentioning

confidence: 99%

Out-of-Distribution Detection of Human Activity Recognition with Smartwatch Inertial Sensors

Boyer

Burns

Whyne

2021

Sensors

View full text Add to dashboard Cite

show abstract

“…These images, referred to as counterfactual-images, are a member of unknown classes but they look like known classes. Using the GAN framework, another work [25] aims at generating synthesis data which were served as fake unknown classes for the classifier to make it robust against real unknown classes. Yu et al [35] proposed the adversarial sample generation (ASG) framework that produces unseen class data.…”

Section: Deep Neural Network-based Algorithmsmentioning

confidence: 99%

A Survey on Open Set Recognition

Mahdavi,

Carvalho

2021

Preprint

View full text Add to dashboard Cite

Recognition (OSR) is about dealing with unknown situations that were not learned by the models during training. In this paper, we provide a survey of existing works about OSR and distinguish their respective advantages and disadvantages to help out new researchers interested in the subject. The categorization of OSR models is provided along with an extensive summary of recent progress. Additionally, the relationships between OSR and its related tasks including multi-class classification and novelty detection are analyzed. It is concluded that OSR can appropriately deal with unknown instances in the real-world where capturing all possible classes in the training data is not practical. Lastly, applications of OSR are highlighted and some new directions for future research topics are suggested.

show abstract

“…The first type includes SVM-based methods [26] and distance-based method [1,2]. A collection of recent OSR works venture towards the generative direction [4,11,19,31]. A subset of OSR methods, named NEL, mainly employ Bayesian methods, such as infinite Gaussian mixture model (IGMM) [24] to learn the UCs.…”

Section: Related Workmentioning

confidence: 99%

“…The presence of such instances weakens the robustness of conventional machine learning algorithms, as these algorithms do not account for the instances from unknown classes, either in the train or the test environments. To overcome this challenge, a series of related research activities has become popular in recent years; examples include anomaly detection (AD) [13,15,27,34], few-shot learning (FSL) [12,25], zero-shot learning (ZSL) [21,29], open set recognition (OSR) and open-world classification (OWC) [1,2,4,5,8,11,14,17,19,20,23,26,30,31]. Collectively, each of these works belongs to one of the four different categories [6], differing on the kind of instances observed by the model during train and test.…”

Section: Introductionmentioning

confidence: 99%

Non-Exhaustive Learning Using Gaussian Mixture Generative Adversarial Networks

Zhuang¹,

Hasan²

2021

Preprint

View full text Add to dashboard Cite

Supervised learning, while deployed in real-life scenarios, often encounters instances of unknown classes. Conventional algorithms for training a supervised learning model do not provide an option to detect such instances, so they miss-classify such instances with 100% probability. Open Set Recognition (OSR) and Non-Exhaustive Learning (NEL) are potential solutions to overcome this problem. Most existing methods of OSR first classify members of existing classes and then identify instances of new classes. However, many of the existing methods of OSR only makes a binary decision, i.e., they only identify the existence of the unknown class. Hence, such methods cannot distinguish test instances belonging to incremental unseen classes. On the other hand, the majority of NEL methods often make a parametric assumption over the data distribution, which either fail to return good results, due to the reason that real-life complex datasets may not follow a well-known data distribution. In this paper, we propose a new online non-exhaustive learning model, namely, Non-Exhaustive Gaussian Mixture Generative Adversarial Networks (NE-GM-GAN) to address these issues. Our proposed model synthesizes Gaussian mixture based latent representation over a deep generative model, such as GAN, for incremental detection of instances of emerging classes in the test data. Extensive experimental results on several benchmark datasets show that NE-GM-GAN significantly outperforms the state-of-the-art methods in detecting instances of novel classes in streaming data.

show abstract

Open Set Recognition by Regularising Classifier with Fake Data Generated by Generative Adversarial Networks

Cited by 36 publications

References 4 publications

Out-of-Distribution Detection of Human Activity Recognition with Smartwatch Inertial Sensors

Out-of-Distribution Detection of Human Activity Recognition with Smartwatch Inertial Sensors

A Survey on Open Set Recognition

Non-Exhaustive Learning Using Gaussian Mixture Generative Adversarial Networks

Contact Info

Product

Resources

About