Learning Less Generalizable Patterns with an Asymmetrically Trained Double Classiﬁer for Better Test-Time Adaptation Thomas Duboudin1Emmanuel Dellandr ea1Corentin Abgrall2Gilles H enaff2Liming Chen1

2025-05-02 0 0 359.63KB 9 页 10玖币

侵权投诉

Learning Less Generalizable Patterns with an Asymmetrically Trained Double

Classiﬁer for Better Test-Time Adaptation

Thomas Duboudin1Emmanuel Dellandr´

ea1Corentin Abgrall2Gilles H´

enaff2Liming Chen1

Univ Lyon, Ecole Centrale de Lyon, CNRS,

INSA Lyon, Univ Claude Bernard Lyon 1,

Univ Louis Lumi`

ere Lyon 2, LIRIS, UMR5205,

69134 Ecully, France

firstname.name@ec-lyon.fr

Thales LAS France,

78990 ´

Elancourt, France

firstname.name@fr.thalesgroup.com

Abstract

Deep neural networks often fail to generalize outside of

their training distribution, in particular when only a single

data domain is available during training. While test-time

adaptation has yielded encouraging results in this setting,

we argue that, to reach further improvements, these ap-

proaches should be combined with training procedure mod-

iﬁcations aiming to learn a more diverse set of patterns. In-

deed, test-time adaptation methods usually have to rely on a

limited representation because of the shortcut learning phe-

nomenon: only a subset of the available predictive patterns

is learned with standard training. In this paper, we ﬁrst

show that the combined use of existing training-time strate-

gies, and test-time batch normalization, a simple adaptation

method, does not always improve upon the test-time adap-

tation alone on the PACS benchmark. Furthermore, ex-

periments on Ofﬁce-Home show that very few training-time

methods improve upon standard training, with or without

test-time batch normalization. We therefore propose a novel

approach using a pair of classiﬁers and a shortcut patterns

avoidance loss that mitigates the shortcut learning behav-

ior by reducing the generalization ability of the secondary

classiﬁer, using the additional shortcut patterns avoidance

loss that encourages the learning of samples speciﬁc pat-

terns. The primary classiﬁer is trained normally, resulting

in the learning of both the natural and the more complex,

less generalizable, features. Our experiments show that our

method improves upon the state-of-the-art results on both

benchmarks and beneﬁts the most to test-time batch nor-

malization.

1. Introduction

Deep neural networks performance falls sharply when

they are confronted, at test-time, with data coming from

a different distribution, or domain, than the training one.

A change in lighting, sensor, weather conditions, or

geographical location can result in a dramatic performance

drop [14, 2, 5]. Such environmental changes are commonly

encountered when an embedded network is deployed in

the wild, and exist in such diversity that it is impossible to

gather enough data to cover all possible domain shifts. This

lack of cross-domain robustness prevents the widespread

deployment of deep networks in safety critical applications.

Domain generalization algorithms have been inves-

tigated to mitigate the test-time performance drop by

modifying the training procedure. Contrary to the domain

adaptation research ﬁeld, in which unlabeled samples of

the target distribution are available at training time [7, 34],

no information about the target domain is assumed to be

known in domain generalization. Most of them assume

to have access to data coming from several identiﬁed

different domains, and try to create a domain invariant

representation by ﬁnding common predictive patterns

[24, 25, 4, 23, 20, 17]. However, such an assumption

is quite generous and in many real-life applications one

does not have access to several data domains but only

a single one. As a result, a number of methods study

single-source domain generalization [38, 31, 44, 43, 27].

A majority of methods were however found to perform

only marginally better than the standard training procedure

when the evaluation is done rigorously on several bench-

marks [9, 42]. Another recent paradigm, called test-time

adaptation, proposes to use a normally trained network,

and adapt it with a quick procedure at test-time, using

arXiv:2210.09834v1 [cs.LG] 17 Oct 2022

only a batch of unlabeled target samples. This paradigm

yielded promising results in the domain generalization

setting [40, 39] because they alleviate the main challenges

of domain generalization: the lack of information about the

target domain, and the requirement to be simultaneously

robust in advance to every possible shift.

Test-time adaptation methods however suffer from a

drawback that limits their adaptation capability and that

can only be corrected at training-time. Indeed, using a

standard training procedure, only a subset of predictive

patterns is learned, corresponding to the most obvious

and efﬁcient ones, while the less predictive patterns are

disregarded entirely [32, 13, 28, 12, 30, 2, 8]. This apparent

ﬂaw, named shortcut learning, originates from the gradient

descent optimization [28] and prevents a test-time method

to use all the available patterns. The combination of a

training-time patterns’ diversity seeking approach with a

test-time adaptation method may thus lead to improved

results. In this paper, we show that the combined use of

test-time batch normalization, a simple test-time adaptation

method, with the state-of-the-art single-source domain

generalization methods (that are often designed to discover

normally unused patterns) does not systematically yield

increased results on the PACS benchmark [22] in the

single-source setting. Similar experiments on Ofﬁce-Home

[35] yield a similar result, with only few methods perform-

ing better than the standard training procedure. We thus

propose a new method, namely L2GP, which encourages

a network to learn new predictive patterns rather than ex-

ploiting and reﬁning already learned ones and demonstrate

its effectiveness on both the PACS and the Ofﬁce-Home

benchmarks. To ﬁnd such patterns, we propose to look for

predictive patterns that are less generalizable than the natu-

rally learned ones through a secondary classiﬁer endowed

with a shortcut avoidance loss, thereby leading to learning

semantically different patterns. These less generalizable

patterns match the ones normally ignored because of the

simplicity bias of deep networks that promotes the learning

of a representation with a high generalization capability

[18, 6]. Our method requires two classiﬁers added to a

features extractor and trains them asymmetrically, using

a data-dependant regularization, e.g., shortcut avoidance

loss, that slightly encourages memorization rather than

generalization by learning batch speciﬁc patterns, i.e.

patterns that lower the loss on the running batch but with a

limited effect on the other batches of data.

To summarize, our contribution is threefold:

• To the best of our knowledge, we are the ﬁrst to inves-

tigate the effect of training-time single-source methods

on a test-time adaptation strategy. We show that it usu-

ally does not increase performance and can even have

Figure 1. Schema of our bi-headed architecture. The naming con-

vention is the same as the one used in algorithm 1.

an adverse effect.

• We apply, for the ﬁrst time, several state-of-the-art

single-source domain generalization algorithms on the

more challenging and rarely used Ofﬁce-Home bench-

mark and show that very few yield a robust cross-

domain representation.

• We propose an original algorithm to learn a larger than

usual subset of predictive features and show that it

yields results over the existing state-of-the-art with the

combination of test-time batch normalization.

2. Related Works

2.1. Single-Source Domain Generalization

Most domain generalization algorithms require several

identiﬁed domains to enforce some level of distributional

invariance. Because this is an unrealistic hypothesis in some

situations (such as in healthcare or defense related tasks),

methods were developed to deal with a domain shift is-

sue with only one single domain available during training.

Some of them rely on a domain shift invariance hypothe-

sis. A commonly used invariance hypothesis is the texture

shift hypothesis: a lot of domain shifts are primarily tex-

tures shifts, and using style transfer based data augmenta-

tion will improve the generalization, whether it is explicitly

by training a model on stylized images [38, 19] or implicitly

in the internal representation of the network [43, 27]. Such

methods are limited to situations where it is indeed a shift of

the hypothesized nature that is encountered. Others wish to

learn a larger set of predictive patterns to make the network

more robust should one or several training-time predictive

patterns be missing at test-time. Volpi et al. [36] and Zhang

et al. [44] propose to incrementally add adversarial images

crafted to maximize the classiﬁcation error of the network

to the training dataset. These images no longer contain the

original obvious predictive patterns which forces the learn-

ing of new patterns. These strategies are inspired by ad-

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：

LearningLessGeneralizablePatternswithanAsymmetricallyTrainedDoubleClassierforBetterTest-TimeAdaptationThomasDuboudin1EmmanuelDellandr´ea1CorentinAbgrall2GillesH´enaff2LimingChen1UnivLyon,EcoleCentraledeLyon,CNRS,INSALyon,UnivClaudeBernardLyon1,UnivLouisLumiereLyon2,LIRIS,UMR5205,69134Ecully,France...

展开>> 收起<<

Learning Less Generalizable Patterns with an Asymmetrically Trained Double Classiﬁer for Better Test-Time Adaptation Thomas Duboudin1Emmanuel Dellandr ea1Corentin Abgrall2Gilles H enaff2Liming Chen1.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！

Learning Less Generalizable Patterns with an Asymmetrically Trained Double Classiﬁer for Better Test-Time Adaptation Thomas Duboudin1Emmanuel Dellandr ea1Corentin Abgrall2Gilles H enaff2Liming Chen1

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: