检索结果-内蒙古大学图书馆

Portable Reward Tuning: Towards Reusable Fine-Tuning across Different Pretrained Models

学校读者我要写书评

暂无评论

arXiv 2025年

作者： Chijiwa, Daiki Hasegawa, Taku Nishida, Kyosuke Saito, Kuniko Takeuchi, Susumu NTT Computer and Data Science Laboratories NTT Corporation Japan NTT Human Informatics Laboratories NTT Corporation Japan

While foundation models have been exploited for various expert tasks through fine-tuning, any foundation model will become outdated due to its old knowledge or limited capability. Thus the underlying foundation model should be eventually replaced by new ones, which leads to repeated cost of fine-tuning these new models. Existing work addresses this problem by inference-time tuning, i.e., modifying the output probabilities from the new foundation model with the outputs from the old foundation model and its fine-tuned model, which involves an additional overhead in inference by the latter two models. In this paper, we propose a new fine-tuning principle, Portable Reward Tuning (PRT), that reduces the inference overhead by its nature, based on the reformulation of fine-tuning as the reward maximization. Specifically, instead of fine-tuning parameters of the foundation models, PRT trains the reward model explicitly through the same loss function as in fine-tuning. During inference, the reward model can be used with any foundation model (with the same set of vocabularies or labels) through the formulation of reward maximization. Experimental results, covering both vision and language models, demonstrate that the PRT-trained model can achieve comparable accuracy to the existing work of inference-time tuning, with less inference cost. Copyright © 2025, The Authors. All rights reserved.

关键词： Reusability

Planning Cyclic Queuing and Forwarding for Time-sensitive Networking via Heuristic and Neurodynamic Optimization

学校读者我要写书评

暂无评论

IEEE Transactions on Network science and Engineering 2025年

作者： Wang, Jiasen Wang, Jun Purple Mountain Laboratories Future Network Research Center Nanjing211111 China City University of Hong Kong Departments of Computer Science and Data Science Hong Kong

Time-sensitive networking (TSN) is a potential means to provide high-quality communication for cyber-physical systems in industrial automation and manufacturing. IEEE 802.1Qch cyclic queuing and forwarding (CQF) is a mechanism for transmitting time-sensitive packets in TSN systems. This paper addresses the planning of periodic time-sensitive flows for CQF in TSN systems. An integer programming problem is formulated to plan the periodic time-sensitive flows. In view of the NP-hardness of the integer programming problem, a polynomial-time heuristic algorithm is developed for CQF planning based on flow efficiency sorting and packet swapping to compute time offsets of multi-period flows. The heuristic algorithm is proven to converge to a CQF plan with a constant worst-case performance ratio with respect to the global optimal plan under given assumptions and derived sufficient conditions. Based on the heuristic solution, neurodynamic optimization is used to search for global optimal plans. Experimental and simulation results on open-loop network flow planning and closed-loop networked model predictive control are elaborated to substantiate the efficacy of the planning algorithms. © 2013 IEEE.

关键词： collaborative neurodynamic optimization Cyclic queuing and forwarding heuristic planning time-sensitive networking

Deep Transfer Learning for Multiclass Dermatology Classification: A Comprehensive Study 3rd

学校读者我要写书评

暂无评论

Deep Transfer Learning for Multiclass Dermatology Classifica...

3rd International Conference on Emerging Electronics and Automation, E2A 2023

作者： Rao, S. V. Achuta Murty, P. T. Satyanarayana Maram, Balajee Garg, Ankit Daniya, T. Kumar, B. Santhosh Data Science Research Laboratories Sree Dattha Institute of Engineering and Science Sheriguda Telangana Hyderabad501 510 India Shri Vishnu Engineering College for Women Bhimavaram India AIT- Department of Computer Science and Engineering Chandigarh University Punjab Gharuan Mohali140055 India GMR Institute of Technology Rajam532127 India Telangana Ibrahimpatnam501506 India

ISBN: (纸本)9789819768011

Accurate and prompt diagnosis is crucial to the treatment of patients with dermatological diseases, which are a significant cause for worry in the medical community. By automating the diagnosing process, deep learning algorithms have recently demonstrated significant promise in the field of dermatology. In-depth research on the use of deep transfer learning for multiclass dermatological classification is presented in this publication. To train deep neural networks, the study makes use of a broad and comprehensive dermatology dataset that includes a variety of skin disorders. Modern pre-trained models are used with transfer learning techniques to make it easier to extract discriminative features. In order to improve classification performance, we delve into the complex process of model fine-tuning and investigate data augmentation techniques. Our research shows that multiclass dermatology classification accuracy has significantly improved, with models delivering reliable findings for a variety of skin disorders. We offer a thorough evaluation of the model's performance, emphasizing its ability to distinguish between disorders that appear to be visually identical and assisting dermatologists in their diagnostic procedures. The field of dermatology will be significantly impacted by the study's findings. Our work helps to the creation of precise, effective, and available dermatological diagnostic tools by bridging the gap between conventional diagnosis techniques and cutting-edge deep learning technologies. We also go over potential difficulties, ethical issues, and prospective future research routes, highlighting how crucial it is to keep this field innovative. This thorough study highlights the importance of artificial intelligence in improving patient care in the field of dermatology and the possibility of deep transfer learning to revolutionize dermatology diagnosis. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

关键词： Medical imaging

Optimal cavity design for minimizing errors in cavity-QED-based atom-photon entangling gates with finite temporal duration

学校读者我要写书评

暂无评论

Physical Review A 2025年第1期111卷 L011701-L011701页

作者： Takeru Utsugi Rui Asaoka Yuuki Tokunaga Takao Aoki Department of Applied Physics Waseda University 3-4-1 Okubo Shinjuku-ku Tokyo 169-8555 Japan Computer & Data Science Laboratories NTT Corporation Musashino Tokyo 180-8585 Japan

We investigate atom-photon entangling gates based on cavity quantum electrodynamics (QED) for a finite photon-pulse duration, where not only the photon loss but also the temporal mode-mismatch of the photon pulse becomes a severe source of error. We analytically derive relations between cavity parameters, including transmittance, length, and effective cross-sectional area of the cavity, that minimize both the photon loss probability and the error rate due to temporal mode-mismatch by taking it into account as state-dependent pulse delay. We also investigate the effects of pulse distortion using numerical simulations for the case of short pulse duration. We believe that these analyses are the first to suggest that a cavity has an optimal length for the atom-photon gate, providing a fundamental guideline for implementing quantum information processing.

关键词： Atoms

Iterative Feature Space Optimization through Incremental Adaptive Evaluation

学校读者我要写书评

暂无评论

arXiv 2025年

作者： Wu, Yanping Huang, Yanyong Chen, Zhengzhang Yao, Zijun Fu, Yanjie Liu, Kunpeng Luo, Xiao Wang, Dongjie University of Kansas Department of Electrical Engineering and Computer Science United States Southwestern University of Finance and Economics Department of Statistics China NEC Laboratories America Data Science and System Security United States Arizona State University Department of Computing and Augmented Intelligence United States Portland State University Department of Computer Science United States University of California Los Angeles Department of Computer Science United States

Iterative feature space optimization involves systematically evaluating and adjusting the feature space to improve downstream task performance. However, existing works suffer from three key limitations: 1) overlooking differences among data samples leads to evaluation bias;2) tailoring feature spaces to specific machine learning models results in overfitting and poor generalization;3) requiring the evaluator to be retrained from scratch during each optimization iteration significantly reduces the overall efficiency of the optimization process. To bridge these gaps, we propose a gEneralized Adaptive feature Space Evaluator (EASE) to efficiently produce optimal and generalized feature spaces. This framework consists of two key components: Feature-Sample Subspace Generator and Contextual Attention Evaluator. The first component aims to decouple the information distribution within the feature space to mitigate evaluation bias. To achieve this, we first identify features most relevant to prediction tasks and samples most challenging for evaluation based on feedback from the subsequent evaluator. These identified feature and samples are then used to construct feature subspaces for next optimization iteration. This decoupling strategy makes the evaluator consistently target the most challenging aspects of the feature space. The second component intends to incrementally capture evolving patterns of the feature space for efficient evaluation. We propose a weighted-sharing multi-head attention mechanism to encode key characteristics of the feature space into an embedding vector for evaluation. Moreover, the evaluator is updated incrementally, retaining prior evaluation knowledge while incorporating new insights, as consecutive feature spaces during the optimization process share partial information. Extensive experiments on fourteen real-world datasets demonstrate the effectiveness of the proposed framework. Our code and data are publicly available Copyright © 2025, The Autho

关键词： Vector spaces

Quantum variational autoencoder utilizing regularized mixed-state latent representations

学校读者我要写书评

暂无评论

Physical Review A 2025年第4期111卷 042416-042416页

作者： Gaoyuan Wang Jonathan Warrell Prashant S. Emani Mark Gerstein Program in Computational Biology and Bioinformatics Yale University New Haven Connecticut 06520 USA Department of Molecular Biophysics and Biochemistry Yale University New Haven Connecticut 06520 USA NEC Laboratories America Princeton New Jersey 08540 USA Department of Computer Science Yale University New Haven Connecticut 06520 USA Department of Statistics & Data Science Yale University New Haven Connecticut 06520 USA Department of Biomedical Informatics & Data Science Yale University New Haven Connecticut 06520 USA

A major challenge in near-term quantum computing is its application to large real-world datasets due to scarce quantum hardware resources. One approach to enabling tractable quantum models for such datasets involves finding low-dimensional representations that preserve essential information for downstream analysis. In classical machine learning, variational autoencoders (VAEs) facilitate efficient data compression, representation learning for subsequent tasks, and novel data generation. However, no quantum model has been proposed that exactly captures all of these features for direct application to quantum data on quantum computers. Some existing quantum models for data compression lack regularization of latent representations, thus preventing direct use for generation and control of generalization. Others are hybrid models with only some internal quantum components, impeding direct training on quantum data. To address this, we present a fully quantum framework, ζ-QVAE, which encompasses all the capabilities of classical VAEs and can be directly applied to map both classical and quantum data to a lower-dimensional space, while effectively reconstructing much of the original state from it. Our model utilizes regularized mixed states to attain optimal latent representations. It accommodates various divergences for reconstruction and regularization. Furthermore, by accommodating mixed states at every stage, it can utilize the full data density matrix and allow for a training objective defined on probabilistic mixtures of input data. Doing so, in turn, makes efficient optimization possible and has potential implications for private and federated learning. In addition to exploring the theoretical properties of ζ-QVAE, we demonstrate its performance on representative genomics and synthetic data. Our results indicate that ζ-QVAE consistently learns representations that better utilize the capacity of the latent space and exhibits similar or better performance compared with

关键词： Quantum computers

Quantized Analog Beamforming Enabled Multi-task Federated Learning Over-the-air

学校读者我要写书评

暂无评论

arXiv 2025年

作者： Yao, Jiacheng Xu, Wei Zhu, Guangxu Yang, Zhaohui Huang, Kaibin Niyato, Dusit National Mobile Communications Research Laboratory Southeast University Nanjing China Purple Mountain Laboratories Nanjing China Shenzhen Research Institute of Big Data Shenzhen China College of Information Science and Electronic Engineering Zhejiang University Hangzhou China Department of Electrical and Electronic Engineering The University of Hong Kong Hong Kong School of Computer Science and Engineering Nanyang Technological University Singapore

Over-the-air computation (AirComp) has recently emerged as a pivotal technique for communication-efficient federated learning (FL) in resource-constrained wireless networks. Though AirComp leverages the superposition property of multiple access channels for computation, it inherently limits its ability to manage inter-task interference in multi-task computing. In this paper, we propose a quantized analog beamforming scheme at the receiver to enable simultaneous multi-task FL. Specifically, inspiring by the favorable propagation and channel hardening properties of large-scale antenna arrays, a targeted analog beamforming method in closed form is proposed for statistical interference elimination. Analytical results reveal that the interference power vanishes by an order of O (1/Nr) with the number of analog phase shifters, Nr, irrespective of their quantization precision. Numerical results demonstrate the effectiveness of the proposed analog beamforming method and show that the performance upper bound of ideal learning without errors can be achieved by increasing the number of low-precision analog phase shifters. Copyright © 2025, The Authors. All rights reserved.

关键词： Multi-task learning

High-purity single-photon generation based on cavity QED

学校读者我要写书评

暂无评论

Physical Review Research 2025年第1期7卷 013251-013251页

作者： Seigo Kikura Rui Asaoka Masato Koashi Yuuki Tokunaga Department of Applied Physics Graduate School of Engineering The University of Tokyo 7-3-1 Hongo Bunkyo-ku Tokyo 113-8656 Japan Faculty of Science and Engineering Waseda University 3-4-1 Okubo Shinjuku-ku Tokyo 169-8585 Japan Computer and Data Science Laboratories NTT Corporation Musashino 180-8585 Japan Photon Science Center Graduate School of Engineering The University of Tokyo 7-3-1 Hongo Bunkyo-ku Tokyo 113-8656 Japan

We propose a scheme for generating a high-purity single photon on the basis of cavity QED. This scheme employs an atom as a four-level system and the structure allows the suppression of the reexcitation process due to the atomic decay, which is known to significantly degrade the single-photon purity in state-of-the-art photon sources using a three-level system. Our analysis shows that the reexcitation probability arbitrarily approaches zero without sacrificing the photon generation probability when increasing the power of a driving laser between two excited states. This advantage is achievable by using current cavity-QED technologies. Our scheme can contribute to developing distributed quantum computation or quantum communication with high accuracy.

关键词： Quantum computers

data-Efficient Error Mitigation for Physical and Algorithmic Errors in a Hamiltonian Simulation

学校读者我要写书评

暂无评论

arXiv 2025年

作者： Hakkaku, Shigeo Suzuki, Yasunari Tokunaga, Yuuki Endo, Suguru NTT Computer and Data Science Laboratories NTT Corporation 3-9-11 Midori-cho Musashino Tokyo180-8585 Japan NTT Research Center for Theoretical Quantum Physics NTT Corporation 3-1 Morinosato-Wakamiya Atsugi243-0198 Japan JST PRESTO 4-1-8 Honcho Saitama Kawaguchi332-0012 Japan

Quantum dynamics simulation via Hamilton simulation algorithms is one of the most crucial applications in the quantum computing field. While this task has been relatively considered the target in the fault-tolerance era, the experiment for demonstrating utility by an IBM team simulates the dynamics of an Ising-type quantum system with the Trotter-based Hamiltonian simulation algorithm with the help of quantum error mitigation [1]. In this study, we propose the data-efficient 1D extrapolation method to mitigate not only physical errors but also algorithmic errors of Trotterized quantum circuits in both the near-term and early fault-tolerant eras. Our proposed extrapolation method uses expectation values obtained by Trotterized circuits, where the Trotter number is selected to minimize both physical and algorithmic errors according to the circuit’s physical error rate. We also propose a method that combines the data-efficient 1D extrapolation with purification QEM methods, which improves accuracy more at the expense of multiple copies of quantum states or the depth of the quantum circuit. Using the 1D transverse-field Ising model, we numerically demonstrate our proposed methods and confirm that our proposed extrapolation method suppresses both statistical and systematic errors more than the previous extrapolation method. © 2025, CC0.

关键词： Ising model