Semi-Supervised Domain Adaptation with Auto-Encoder via Simultaneous Learning Md Mahmudur Rahman

2025-05-03 1 0 1.12MB 10 页 10玖币

侵权投诉

Semi-Supervised Domain Adaptation with Auto-Encoder via Simultaneous

Learning

Md Mahmudur Rahman

University of Massachusetts Lowell

mdmahmudur rahman@student.uml.edu

Rameswar Panda

MIT-IBM Watson AI Lab

rpanda@ibm.com

Mohammad Arif Ul Alam

University of Massachusetts Lowell

mohammadariful alam@uml.edu

Abstract

We present a new semi-supervised domain adapta-

tion framework that combines a novel auto-encoder-based

domain adaptation model with a simultaneous learning

scheme providing stable improvements over state-of-the-art

domain adaptation models. Our framework holds strong

distribution matching property by training both source and

target auto-encoders using a novel simultaneous learn-

ing scheme on a single graph with an optimally modiﬁed

MMD loss objective function. Additionally, we design a

semi-supervised classiﬁcation approach by transferring the

aligned domain invariant feature spaces from source do-

main to the target domain. We evaluate on three datasets

and show proof that our framework can effectively solve

both fragile convergence (adversarial) and weak distribu-

tion matching problems between source and target feature

space (discrepancy) with a high ‘speed’ of adaptation re-

quiring a very low number of iterations.

1. Introduction

Deep domain adaptation has become an emerging learning

technique to address the lack of labeled data with the help

of data from related but different domains[31, 5, 31]. The

key advantages of deep domain adaptation networks are the

ability to train models from partially labeled or even unla-

beled data with the help of related source domain by trans-

ferring knowledge among those domains. Much progress

has been done in deep domain adaptation focusing on Un-

supervised Domain Adaptation (UDA) or Semi-supervised

domain Adaptation(SSDA). Although UDA methods can

be converted to SSDA methods by introducing supervision

from labeled target samples, both methods cannot guaran-

tee the inter-domain feature distribution matching, espe-

Figure 1: Illustration of advantages of using our model. Here solid

circle, dashed circle, solid line (red for before DA/green for after

DA) represent source, target and decision boundary respectively.

(a) decision boundary violates the feature space before adaptation.

(b) While training with our framework iteratively, both of the fea-

ture spaces align each other. (c) Semi-supervised ﬁne-tuning en-

ables the decision boundaries to structure according to the newly

adapted source and target feature spaces. Finally, (d) our method

generates a robust decision boundary.

cially with the unlabeled samples [16] due to the presence of

intra-domain feature space discrepancy resulting poor per-

formance [27].

Based on the feature space mapping techniques, semi-

supervised domain adaptation can be divided into three ma-

jor categories: Discrepancy-based, adversarial discrimina-

tive, and adversarial generative models [5]. In adversar-

ial discriminative models, a domain discriminator network

tries to encourage domain confusion by optimizing an ad-

arXiv:2210.09486v1 [cs.CV] 18 Oct 2022

versarial loss function [31]. A generative component is at-

tached to the discriminative network to generate more sim-

ilar feature spaces for source and target domains in the

adversarial generative models [10]. Although adversarial

models achieve signiﬁcant improvement, the training pro-

cess of adversarial networks is complicated that often re-

sults very fragile convergence [1]. Unlike the adversarial

networks, a shared feature space has been learned by min-

imizing the distribution discrepancy between source and

target domains in discrepancy-based methods [24, 39, 29].

The main advantage of discrepancy-based methods is they

are easy to train and convergence is not as fragile as ad-

versarial networks [6]. However, most discrepancy-based

methods minimize the discrepancy between source and tar-

get domains and very few of them address discrepancy

within domains [40, 14, 35]. Especially when there is a

low correlation between labeled and unlabeled samples in

the target domain which results only partial alignment with

the target feature space [35]. As a result, discrepancy-based

methods can not guarantee distribution matching and often

suffer from over-ﬁtting to the target domain[2].

In this paper, we introduce a novel training concept

called simultaneous learning to address above shortcom-

ings. Aligning to a ﬁxed source feature space can be prob-

lematic because of the presence of an intra-domain discrep-

ancy within the target domain [16]. To address this is-

sue, we train both source and target network simultaneously

where both feature spaces align one another in the course of

learning (Figure 1). We implement simultaneous learning

with an auto-encoder-based domain adaptation where fea-

ture spaces of source and target domain get aligned in the

bottleneck layer. In this framework, ﬁrst, the simultane-

ous learning scheme allows the source domain to align with

the target domain along with the intra-domain discrepan-

cies present in the target domain. Concurrently, the target

domain also gets aligned with the source domain with its

intra-domain discrepancies. As a result, after learning iter-

atively, both source and target domain feature spaces align

with each other including their intra-domain discrepancies,

and solve the problem of partial alignment. Second, we use

the bottleneck layer of auto-encoder as feature space rather

than the layers in the classiﬁcation network used in other

methods. As the learning objective of the auto-encoder is

not to classify samples, this enables a smooth transition and

continuous feature space among the classes. If any feature

discrepancy presents within the classes, continuous feature

spaces help them to align completely with the other domain.

To best of our knowledge, our proposed method is a ﬁrst-

of-its-kind framework that considers a simultaneous learn-

ing scheme with source and target networks with a carefully

modiﬁed MMD loss across the domain invariant feature

space. Our proposed framework is both easy to train and

achieve state-of-the-art performance by solving the problem

of partial alignment. key contributions:

•Propose a novel simultaneous learning scheme to help tar-

get feature space to align with the source domain feature

space for SSDA.

•Modify MMD loss function in a way which works with si-

multaneous learning and achieve strong inter-domain align-

ment.

•Propose an auto-encoder-based SSDA framework to ad-

dress intra-domain discrepancy issue by implementing the

simultaneous learning scheme and evaluate performances

with state-of-the-art models using three datasets .

2. Related Works

Unsupervised domain adaptation (UDA) has become

mostly popular recently that can be categorized in three ap-

proaches. First one is to reduce the distribution discrep-

ancy by adversarial learning [4] [36] [7] [37] which has

few variances, such as utilization of adversarial learning

to introduce an intermediate domain approach with Grad-

ual Vanishing Bridge (GVB) layer between source and tar-

get domain [7]. The second type of method minimizes the

domain divergence with different loss functions, such as

MMD [24], CORAL [30], HoMM [2] etc. Recently, Do-

main Conditioned Adaptation Network (DCAN) [21] uses

domain conditioned channel attention mechanism to iden-

tify the domain speciﬁc features. The last type of method

uses optimal transport (OT) mechanism which minimizes

the cost of transporting one feature distribution to another

domain [3] [8][38][20]. Semi-supervised domain adapta-

tion (SSDA) fuses the well-labeled source domain with the

label scarce target domain with the help of few labeled tar-

get samples. Kim et al. [16] introduce the intra-domain

discrepancy between the labeled and unlabeled data in the

target domain and minimize it to align source and target do-

main features. Another approach uses meta-learning frame-

work for multi-source and semi-supervised domain adap-

tation [19]. BiAT framework uses bidirectional adversarial

training to guide the adversarial examples across the domain

gap and achieve domain alignment [15]. Another adversar-

ial learning based approach MME uses conditional minmax

entropy optimization to achieve adaptation [27].

3. Proposed Method

Let the labeled source domain is Ds={(Xs

i, Y s

i)}ns.

Here Xs

i={x1, ..., xns} ∈ Rdsis the ith sample from the

source image with a feature space of dsdimension. Simi-

larly, Ys

i∈ Ysis the class label of corresponding ith sam-

ple of the source images Xs

i, where Ys={1,2, ..., nc}.

Now, for our semi-supervised setup let consider the target

domain Dtas union of two subset, labeled target domain,

Dland unlabeled target domain Du. So, Dt=Dl∪ Du=

{(Xl

i, Y l

i)}nl∪{(Xu

i}nuwhere Xl

i∈Rdtis the labeled ith

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：

Semi-SupervisedDomainAdaptationwithAuto-EncoderviaSimultaneousLearningMdMahmudurRahmanUniversityofMassachusettsLowellmdmahmudurrahman@student.uml.eduRameswarPandaMIT-IBMWatsonAILabrpanda@ibm.comMohammadArifUlAlamUniversityofMassachusettsLowellmohammadarifulalam@uml.eduAbstractWepresentanewsemi-super...

展开>> 收起<<

Semi-Supervised Domain Adaptation with Auto-Encoder via Simultaneous Learning Md Mahmudur Rahman.pdf

共10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！

Semi-Supervised Domain Adaptation with Auto-Encoder via Simultaneous Learning Md Mahmudur Rahman

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: