PQLM - MULTILINGUAL DECENTRALIZED PORTABLE QUANTUM LANGUAGE MODEL Shuyue Stella Li1Xiangyu Zhang1Shu Zhou3Hongchao Shu1 Ruixing Liang1Hexin Liu4Leibny Paola Garcia12

2025-05-02 1 0 2.42MB 5 页 10玖币

侵权投诉

PQLM - MULTILINGUAL DECENTRALIZED PORTABLE QUANTUM LANGUAGE MODEL

Shuyue Stella Li?1Xiangyu Zhang?1Shu Zhou3Hongchao Shu1

Ruixing Liang1Hexin Liu4Leibny Paola Garcia1,2

1Center for Language and Speech Processing, Johns Hopkins University

2Human Language Technology Center of Excellence, Johns Hopkins University

3Department of Physics, Hong Kong University of Science and Technology

4School of Electrical and Electronic Engineering, Nanyang Technological University

ABSTRACT

With careful manipulation, malicious agents can reverse en-

gineer private information encoded in pre-trained language

models. Security concerns motivate the development of

quantum pre-training. In this work, we propose a highly

portable quantum language model (PQLM) that can easily

transmit information to downstream tasks on classical ma-

chines. The framework consists of a cloud PQLM built with

random Variational Quantum Classiﬁers (VQC) and local

models for downstream applications. We demonstrate the ad

hoc portability of the quantum model by extracting only the

word embeddings and effectively applying them to down-

stream tasks on classical machines. Our PQLM exhibits

comparable performance to its classical counterpart on both

intrinsic evaluation (loss, perplexity) and extrinsic evaluation

(multilingual sentiment analysis accuracy) metrics. We also

perform ablation studies on the factors affecting PQLM per-

formance to analyze model stability. Our work establishes a

theoretical foundation for a portable quantum pre-trained lan-

guage model that could be trained on private data and made

available for public use with privacy protection guarantees.

Index Terms—Quantum Machine Learning, Language

Modeling, Federated Learning, Model Portability

1. INTRODUCTION

A competitive language model can be extremely useful for

downstream tasks such as machine translation and speech

recognition despite the domain mismatch between pre-training

and downstream tasks [1, 2]. They become more powerful

with increased training data, but there is a trade-off between

data privacy and utility[3]. Previous works on ethical AI have

shown that pre-trained language models (PLM)s memorizes

training data in addition to learning about the language [4, 5],

which opens up vulnerabilities for potential adversaries to

recover sensitive training data from the model.

While some argue that language models should only be

trained on data explicitly produced for public use [6], pri-

?Equal contribution in alphabetical order

Fig. 1: Decentralized Quantum Language Model Pipeline.

Text data is trained on language model on NISQ servers, the

word embeddings are transferred to downstream models Mi

vate data are richer in certain domains compared to public

corpora, including dialogue systems, code-mixing languages,

and medical applications [7, 8]. Therefore, it is essential to

develop new methods to mitigate potential data security and

privacy problems while being able to take advantage of the

rich linguistic information encoded in private data.

Recently, there has been growing interest in leveraging

random quantum circuits in neural models to solve data pri-

vacy issues [9]. The entanglement of states from the random

conﬁguration of gates in the quantum circuits makes it pos-

sible to securely encode sensitive information contained in

training data [10, 9]. The combination of random quantum

circuits and decentralized training ensures privacy [11]. Ad-

ditionally, quantum computing has become the next logical

step in the development of deep learning for its efﬁciency in

manipulating large tensors [12].

The architecture of a large quantum computer is vastly

different from the classical computer as physical and environ-

mental constraints cannot be met [13], which means that the

model trained on the large quantum computer is difﬁcult to

be directly used by others on the classical computer. Ad hoc

portability is deﬁned as the model’s ability to transmit the

most essential information contained in the language model

arXiv:2210.03221v5 [cs.LG] 27 Feb 2023

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：

PQLM-MULTILINGUALDECENTRALIZEDPORTABLEQUANTUMLANGUAGEMODELShuyueStellaLi?1XiangyuZhang?1ShuZhou3HongchaoShu1RuixingLiang1HexinLiu4LeibnyPaolaGarcia1,21CenterforLanguageandSpeechProcessing,JohnsHopkinsUniversity2HumanLanguageTechnologyCenterofExcellence,JohnsHopkinsUniversity3DepartmentofPhysics,Hong...

展开>> 收起<<

PQLM - MULTILINGUAL DECENTRALIZED PORTABLE QUANTUM LANGUAGE MODEL Shuyue Stella Li1Xiangyu Zhang1Shu Zhou3Hongchao Shu1 Ruixing Liang1Hexin Liu4Leibny Paola Garcia12.pdf

共5页,预览1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！

PQLM - MULTILINGUAL DECENTRALIZED PORTABLE QUANTUM LANGUAGE MODEL Shuyue Stella Li1Xiangyu Zhang1Shu Zhou3Hongchao Shu1 Ruixing Liang1Hexin Liu4Leibny Paola Garcia12

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: