CONNECTIVE RECONSTRUCTION -BASED NOVELTY DETECTION Seyyed Morteza Hashemi Institute for Advanced Studies in Basic Sciences IASBS

2025-05-02 0 0 2.14MB 9 页 10玖币

侵权投诉

CONNECTIVE RECONSTRUCTION-BASED NOVELTY DETECTION

Seyyed Morteza Hashemi

Institute for Advanced Studies in Basic Sciences (IASBS)

MortezaHashemi@iasbs.ac.ir

Parvaneh Aliniya

University of Nevada, Reno

Aliniya@nevada.unr.edu

Parvin Razzaghi

Institute for Advanced Studies in Basic Sciences (IASBS)

P.razzaghi@iasbs.ac.ir

ABSTRACT

Detection of out-of-distribution samples is one of the critical tasks for real-world applications of

computer vision. The advancement of deep learning has enabled us to analyze real-world data which

contain unexplained samples, accentuating the need to detect out-of-distribution instances more

than before. GAN-based approaches have been widely used to address this problem due to their

ability to perform distribution ﬁtting; however, they are accompanied by training instability and

mode collapse. We propose a simple yet efﬁcient reconstruction-based method that avoids adding

complexities to compensate for the limitations of GAN models while outperforming them. Unlike

previous reconstruction-based works that only utilize reconstruction error or generated samples, our

proposed method simultaneously incorporates both of them in the detection task. Our model, which

we call "Connective Novelty Detection" has two subnetworks, an autoencoder, and a binary classiﬁer.

The autoencoder learns the representation of the positive class by reconstructing them. Then, the

model creates negative and connected positive examples using real and generated samples. Negative

instances are generated via manipulating the real data, so their distribution is close to the positive

class to achieve a more accurate boundary for the classiﬁer. To boost the robustness of the detection

to reconstruction error, connected positive samples are created by combining the real and generated

samples. Finally, the binary classiﬁer is trained using connected positive and negative examples.

We demonstrate a considerable improvement in novelty detection over state-of-the-art methods on

MNIST and Caltech-256 datasets.

1 Introduction

Novelty detection identiﬁes whether a given sample is out-of-distribution or within the training data distribution.

Out-of-distribution instances are referred to as novel instances (negative), which the method does not have access

to them during the training. It is closely related to topics such as anomaly detection, outlier detection, and open set

recognition. Thanks to the emergence of the powerful computational system and the ever-progressing deep learning

approaches, the research community is recently moving forward to designing approaches to leverage real-world and

online datasets rather than predeﬁned ones. This, in turn, has brought novelty detection into the spotlight in the recent

decade because one of the main challenges of using real-world datasets is that they contain many uncaptured or poorly

represented samples.

As a result, models will not be able to learn the accurate distribution of the training data (positive class). Therefore

when presented with new examples, they will fail to determine whether this unexplained example is from the training

data distribution.

One common way to evaluate the performance of a novelty detection model is to select a category as a positive class

for training a classiﬁer, and in the test stage, use a combination of samples from this class and one or several other

classes (negative classes). In this way, the performance of the model demonstrates its ability to distinguish new samples

within the positive class from others. Having only access to a positive set in training means that the performance

arXiv:2210.13917v1 [cs.CV] 25 Oct 2022

Figure 1: The t-SNE visualization of the positive, generated hard negative and out-of-distribution samples for

the MNIST dataset. Distributions of generated hard negative examples are very close to those of positive samples.

With proper training of the classiﬁer using positive and generated negative samples, the classiﬁer learns the accurate

boundaries to distinguish between positive and out-of-distribution samples. As a result, out-of-distribution samples can

be recognized with a higher probability in the evaluation phase.

of a model is closely related to its ability to learn the representation of the positive class. One approach is using a

one-class classiﬁer that learns the representation of the positive class such as one class SVM [

] and GAN-based

methods [

]. In the test phase, query samples with low classiﬁcation scores for the positive class will be labeled as

negative. Another interesting direction is leveraging autoencoder to reconstruct the training examples in a GAN-based

model [

]. In this methods, the reconstructor tries to generate an authentic version of the input, and the

classiﬁer attempts to distinguish between actual and reconstructed samples. A critical property of this variation of a

one-class classiﬁer is that as the reconstructor is trained using positive samples, it will destroy negative examples in the

test phase, making it easier for the classiﬁer to discriminate between two classes.

Our proposed model is close to the latter approach in terms of using reconstruction to make the negative samples more

distinguishable from the positive ones; however, there are several downsides that we addressed in a new formulation

of the problem by using a binary classiﬁer in a non-GAN setting. The motivation behind this modiﬁcation comes

from the fact that the signiﬁcant performance improvement of these approaches comes from destroying the negative

samples in the reconstruction phase. In addition, training a GAN model is entangled with speciﬁc challenges. In order

to formulate the method as a GAN model, the input of a classiﬁer will be reconstructed image, limiting learning the

representation of the positive class to the accurateness of generated images. Therefore we model a reconstructor as a

denoising autoencoder (Unet [

]) by customizing vanilla Unet to the task in a way that this new version improves the

reconstruction process for the positive samples and deteriorates it for negative ones. Also, the denoising version makes

it robust to noise. In the case of the traditional one-class classiﬁer, the model only distinguishes the sample of two

classes based on their similarity to the learned positive class’ characteristics, as it is only trained on positive samples.

However, if the model has access to the negative class, it would develop a more accurate classiﬁcation boundary and

classify a test image by considering the learned properties of both classes. We introduce a procedure to create negative

samples via manipulating the positive samples. While forcing them to differentiate from the positive instances, the

proposed method keeps negative samples close enough to positive ones in order to create hard negative examples,

enabling the model to form a more accurate boundary between the classes. Therefore we leverage a binary classiﬁer

that is trained on two positive and generated negative examples.

Negative samples are created by applying data augmentation techniques to the reconstructed positive samples.To make

the negative examples “hard”, the resulting image is combined with the original one by a weighted summation, creating

“connective negative examples”. As the reconstructor is trained separately, the reconstruction error for the positive

examples might propagate to training the classiﬁer; reconstructed positive samples are also combined with original

images, in the same way, forming “connective positive samples” to mitigate error propagation. By creating connective

positive and negative examples, we also incorporate reconstruction error in the training of the classiﬁer. For positive

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

CONNECTIVERECONSTRUCTION-BASEDNOVELTYDETECTIONSeyyedMortezaHashemiInstituteforAdvancedStudiesinBasicSciences(IASBS)MortezaHashemi@iasbs.ac.irParvanehAliniyaUniversityofNevada,RenoAliniya@nevada.unr.eduParvinRazzaghiInstituteforAdvancedStudiesinBasicSciences(IASBS)P.razzaghi@iasbs.ac.irABSTRACTDetect...

展开>> 收起<<

CONNECTIVE RECONSTRUCTION -BASED NOVELTY DETECTION Seyyed Morteza Hashemi Institute for Advanced Studies in Basic Sciences IASBS.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

CONNECTIVE RECONSTRUCTION -BASED NOVELTY DETECTION Seyyed Morteza Hashemi Institute for Advanced Studies in Basic Sciences IASBS

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: