Accurate and reliable nuclear data libraries are essential for calculation and design of advanced nuclea systems. A 1200 fine group nuclear data library Hybrid Evaluated Nuclear Data Library/Fine Group(HENDL/FG with n...Accurate and reliable nuclear data libraries are essential for calculation and design of advanced nuclea systems. A 1200 fine group nuclear data library Hybrid Evaluated Nuclear Data Library/Fine Group(HENDL/FG with neutrons of up to 150 Me V has been developed to improve the accuracy of neutronics calculations and anal ysis. Corrections of Doppler, resonance self-shielding, and thermal upscatter effects were done for HENDL/FG Shielding and critical safety benchmarks were performed to test the accuracy and reliability of the library. The dis crepancy between calculated and measured nuclea parameters fell into a reasonable range.展开更多
基于推理数据的大模型指令微调,通过显式建模复杂任务的多步逻辑关联,显著提升模型的推理准确性,然而微调过程依赖海量高质量数据,导致算力开销急剧攀升。已有数据压缩技术主要聚焦于原始规模缩减,普遍缺乏针对推理数据的压缩方法设计,...基于推理数据的大模型指令微调,通过显式建模复杂任务的多步逻辑关联,显著提升模型的推理准确性,然而微调过程依赖海量高质量数据,导致算力开销急剧攀升。已有数据压缩技术主要聚焦于原始规模缩减,普遍缺乏针对推理数据的压缩方法设计,忽视了推理数据中的多步逻辑关联、语义依存关系等,致使关键推理链完整性受损,进而降低了推理性能。为此,提出基于推理贡献度的精化(Refinement Based on Inference Contribution,RBIC)。通过分析推理数据的语义相似性构建知识领域图谱,精准定位核心信息。将数据样本的语义与大模型的推理准确率相结合,划分难度梯度,覆盖全场景推理需求。通过多步推理数据的逻辑复杂度量化推理贡献度,精化对模型推理贡献度最高的数据样本。实验结果表明,基于RBIC精化的推理数据进行微调后,模型平均推理性能仅下降1.13%,而训练时间缩短为原耗时的16%,验证了RBIC在模型效能与资源消耗间实现了最优平衡,有望推动多领域大模型在资源受限环境下的高效部署与微调优化。展开更多
随着深度学习和神经网络技术的不断发展,光学字符识别(Optical Character Recognition,OCR)技术在从各类图像中准确识别文本信息方面取得了显著进展,无论是印刷体文档、手写笔记,还是包含文字的图片,OCR技术都能较为准确地提取其中的文...随着深度学习和神经网络技术的不断发展,光学字符识别(Optical Character Recognition,OCR)技术在从各类图像中准确识别文本信息方面取得了显著进展,无论是印刷体文档、手写笔记,还是包含文字的图片,OCR技术都能较为准确地提取其中的文字内容,为信息处理和数字化提供了极大的便利。然而在实际应用场景中,OCR技术面临着诸多挑战。不同的光照条件、字体样式、背景干扰等,都会影响识别的准确性。同时,还存在着诸多单点问题,比如特定领域的专业术语识别困难、不规则排版文字的处理不佳等。如何快速响应用户的个性化定制需求,并实现模型的高效微调与快速迭代,已成为影响用户体验的关键因素之一。针对OCR技术在移动端模型大规模数据微调中的实际需求,该文重点探讨了数据清洗与模型微调技术在快速微调实践中的应用与效果。通过有效的数据清洗,可以去除噪声数据,提高数据质量,为模型微调提供更好的基础。而模型微调技术则能够让模型更快地适应新的需求,实现快速迭代,从而提升OCR技术在实际应用中的性能和用户体验。展开更多
服装已经成为网络购物的重要商品之一,实现精准的符合用户个性化审美的服装推荐系统,已经成为热门研究内容。针对提取用户的细粒度兴趣特征不全面,导致推荐系统的准确性低问题,提出融合长短期偏好的服装推荐算法;针对数据稀疏以及数据...服装已经成为网络购物的重要商品之一,实现精准的符合用户个性化审美的服装推荐系统,已经成为热门研究内容。针对提取用户的细粒度兴趣特征不全面,导致推荐系统的准确性低问题,提出融合长短期偏好的服装推荐算法;针对数据稀疏以及数据单一性,导致推荐结果个性化、多样性低的问题,利用跨模态数据和注意力机制使模型学习出更为精准的差异性用户特征。在真实数据集Clothing Shoes and Jewelry上,将所设计的模型(PCR)与经典的循环神经网络RNN、基于矩阵分解MF-BPR模型以及改进的矩阵分解TARMF模型进行性能比对,PCR模型在关键性能评价指标NDCG、Precision@K和Recall@K均有提升。实验结果表明该模型在服装推荐系统中是可行与有效的。展开更多
Background: The Canadian province of Saskatchewan introduced a pre-fine needle aspiration (FNA) clinic to review adherence of referrals for thyroid biopsy based on the guidelines of the American College of Radiology’...Background: The Canadian province of Saskatchewan introduced a pre-fine needle aspiration (FNA) clinic to review adherence of referrals for thyroid biopsy based on the guidelines of the American College of Radiology’s (ACR) Thyroid Imaging, Reporting and Data System (TI-RADS) scoring system. The intention is to minimize low-yield biopsy rates by improving the quality of thyroid nodule investigation in Saskatchewan through this clinic. TI-RADS is a malignancy risk scoring system for thyroid nodules based on five sonographic characteristics: composition, echogenicity, shape, margin, and echogenic foci (calcium). Recommendations for intervention or clinical follow-up are further determined by the size of the nodule. Methods: Through a retrospective chart review of all thyroid biopsy referrals to the Royal University Hospital (RUH) in Saskatchewan between 22 March 2016 and 17 May 2018, the impact of the multidisciplinary pre-FNA clinic on appropriate thyroid biopsies in Saskatchewan was evaluated. Results: This study evaluated 252 referrals, 203 of which underwent FNA and 23 which received surgical biopsy. TI-RADS scores appended to thyroid biopsy referrals increased upon pre-FNA clinic initiation, yet score quality did not improve. Rates of malignant biopsies were lower than ACR-reporting suggesting inappropriate biopsy of low risk nodules perhaps by overcalling the TI-RADS score. The majority of FNA cytology matched final surgical pathology, with 78% of indeterminate FNAs being malignant, and all non-diagnostic FNAs being benign. Conclusions: The implementation of the pre-FNA clinic reduced the number of thyroid biopsies in Saskatchewan by 11% overall.展开更多
基金supported by the Natural Science Foundation of China(Nos.11405204 11305205 and 10675123)Special Program for Informatization of Chinese Academy of Sciences(No.XXH12504-1-09)the National Special Program for ITER(No.2014GB1120001)
文摘Accurate and reliable nuclear data libraries are essential for calculation and design of advanced nuclea systems. A 1200 fine group nuclear data library Hybrid Evaluated Nuclear Data Library/Fine Group(HENDL/FG with neutrons of up to 150 Me V has been developed to improve the accuracy of neutronics calculations and anal ysis. Corrections of Doppler, resonance self-shielding, and thermal upscatter effects were done for HENDL/FG Shielding and critical safety benchmarks were performed to test the accuracy and reliability of the library. The dis crepancy between calculated and measured nuclea parameters fell into a reasonable range.
文摘基于推理数据的大模型指令微调,通过显式建模复杂任务的多步逻辑关联,显著提升模型的推理准确性,然而微调过程依赖海量高质量数据,导致算力开销急剧攀升。已有数据压缩技术主要聚焦于原始规模缩减,普遍缺乏针对推理数据的压缩方法设计,忽视了推理数据中的多步逻辑关联、语义依存关系等,致使关键推理链完整性受损,进而降低了推理性能。为此,提出基于推理贡献度的精化(Refinement Based on Inference Contribution,RBIC)。通过分析推理数据的语义相似性构建知识领域图谱,精准定位核心信息。将数据样本的语义与大模型的推理准确率相结合,划分难度梯度,覆盖全场景推理需求。通过多步推理数据的逻辑复杂度量化推理贡献度,精化对模型推理贡献度最高的数据样本。实验结果表明,基于RBIC精化的推理数据进行微调后,模型平均推理性能仅下降1.13%,而训练时间缩短为原耗时的16%,验证了RBIC在模型效能与资源消耗间实现了最优平衡,有望推动多领域大模型在资源受限环境下的高效部署与微调优化。
文摘随着深度学习和神经网络技术的不断发展,光学字符识别(Optical Character Recognition,OCR)技术在从各类图像中准确识别文本信息方面取得了显著进展,无论是印刷体文档、手写笔记,还是包含文字的图片,OCR技术都能较为准确地提取其中的文字内容,为信息处理和数字化提供了极大的便利。然而在实际应用场景中,OCR技术面临着诸多挑战。不同的光照条件、字体样式、背景干扰等,都会影响识别的准确性。同时,还存在着诸多单点问题,比如特定领域的专业术语识别困难、不规则排版文字的处理不佳等。如何快速响应用户的个性化定制需求,并实现模型的高效微调与快速迭代,已成为影响用户体验的关键因素之一。针对OCR技术在移动端模型大规模数据微调中的实际需求,该文重点探讨了数据清洗与模型微调技术在快速微调实践中的应用与效果。通过有效的数据清洗,可以去除噪声数据,提高数据质量,为模型微调提供更好的基础。而模型微调技术则能够让模型更快地适应新的需求,实现快速迭代,从而提升OCR技术在实际应用中的性能和用户体验。
文摘服装已经成为网络购物的重要商品之一,实现精准的符合用户个性化审美的服装推荐系统,已经成为热门研究内容。针对提取用户的细粒度兴趣特征不全面,导致推荐系统的准确性低问题,提出融合长短期偏好的服装推荐算法;针对数据稀疏以及数据单一性,导致推荐结果个性化、多样性低的问题,利用跨模态数据和注意力机制使模型学习出更为精准的差异性用户特征。在真实数据集Clothing Shoes and Jewelry上,将所设计的模型(PCR)与经典的循环神经网络RNN、基于矩阵分解MF-BPR模型以及改进的矩阵分解TARMF模型进行性能比对,PCR模型在关键性能评价指标NDCG、Precision@K和Recall@K均有提升。实验结果表明该模型在服装推荐系统中是可行与有效的。
文摘Background: The Canadian province of Saskatchewan introduced a pre-fine needle aspiration (FNA) clinic to review adherence of referrals for thyroid biopsy based on the guidelines of the American College of Radiology’s (ACR) Thyroid Imaging, Reporting and Data System (TI-RADS) scoring system. The intention is to minimize low-yield biopsy rates by improving the quality of thyroid nodule investigation in Saskatchewan through this clinic. TI-RADS is a malignancy risk scoring system for thyroid nodules based on five sonographic characteristics: composition, echogenicity, shape, margin, and echogenic foci (calcium). Recommendations for intervention or clinical follow-up are further determined by the size of the nodule. Methods: Through a retrospective chart review of all thyroid biopsy referrals to the Royal University Hospital (RUH) in Saskatchewan between 22 March 2016 and 17 May 2018, the impact of the multidisciplinary pre-FNA clinic on appropriate thyroid biopsies in Saskatchewan was evaluated. Results: This study evaluated 252 referrals, 203 of which underwent FNA and 23 which received surgical biopsy. TI-RADS scores appended to thyroid biopsy referrals increased upon pre-FNA clinic initiation, yet score quality did not improve. Rates of malignant biopsies were lower than ACR-reporting suggesting inappropriate biopsy of low risk nodules perhaps by overcalling the TI-RADS score. The majority of FNA cytology matched final surgical pathology, with 78% of indeterminate FNAs being malignant, and all non-diagnostic FNAs being benign. Conclusions: The implementation of the pre-FNA clinic reduced the number of thyroid biopsies in Saskatchewan by 11% overall.
