id string | image string | input string | system string | output string | boxes list | box_texts list | detail dict |
|---|---|---|---|---|---|---|---|
2010.04529-zh-3 | 2010.04529-zh-3.jpg | <FigureHere>
Figure 1: PolyTope verdicts each error by three coordinates according to its syntactic and semantic roles.
示，与 ROUGE 相比，PolyTope 在提供整体
质量的详细和诊断方面更加精细
我们开发了一个用于注释的操作界面，如附录 A 所示。1。特别是，将原始文本和输出摘要并列显示给人类注释者，并要求他们选择阅读后认为不正确的部分。经过初步选择，他被要求分别在 8 个问题类型和 8 个语法标签中进行进一步选择。然后为每个不正确的段自动生成嵌入的严重性分数，并为注释的摘要计算质量分数，如下所示：
... | [
[
0.298828125,
0.0748291015625,
0.370361328125,
0.07244873046875,
0.370361328125,
0.0843505859375,
0.298828125,
0.08612060546875
],
[
0.52490234375,
0.07781982421875,
0.615234375,
0.07781982421875,
0.615234375,
0.08905029296875,
0.52490234375,
0.08... | [
"Syntactic Label",
"Addition",
"摘要中包含来自源的不必要和不",
"",
"相关的代码段。",
"Severity",
"Aribate",
"Omission输出中缺少关键点。",
"Oesuene",
"Cfies",
"InaccuracyIntrinsic来源的术语或概念被歪",
"曲，因此不忠",
"InaccuracyExtrinsic摘要的内容未在源代",
"码中显示，并且实际上不正确。",
"nand",
"Pos Nae",
"Positive-NegativeAspect输出摘要表示肯",
"Wor_For... | {
"arxiv_id": "2010.04529",
"language": "zh",
"page_idx": 3
} | ||
2010.04529-zh-4 | 2010.04529-zh-4.jpg | <table>
<tbody>
<tr>
<th>type lssue</th>
<th>Sub Issue Type</th>
<th>Subject</th>
<th>Object</th>
<th>Predicate</th>
<th>Number $:\&$Time</th>
<th>Place $\&$ Name</th>
<th>Attribute</th>
<th>Function Word</th>
<th>Whole Sentenco</th>
</tr>
<tr>
<td>$\operatorname{Accuracy}$</td>
... | [
[
0.145263671875,
0.07244873046875,
0.2359619140625,
0.07244873046875,
0.2359619140625,
0.0831298828125,
0.145263671875,
0.0831298828125
],
[
0.55322265625,
0.07244873046875,
0.5693359375,
0.0706787109375,
0.57080078125,
0.07659912109375,
0.55517578125... | [
"Issue typeSub I",
"DT",
"Whole",
"Wor",
"Type",
"At",
"ntenc",
"",
"Addition",
"Critica",
"Major",
"Major",
"Minor",
"Major",
"Critical",
"ica",
"Major",
"Critica",
"Omission",
"Major",
"Minor",
"Critical",
"Critical",
"Critical",
"",
"",
"绘",
"Critical",
"Cr... | {
"arxiv_id": "2010.04529",
"language": "zh",
"page_idx": 4
} | ||
2206.10801-zh-0 | 2206.10801-zh-0.jpg | # Automated Cancer Subtyping via Vector Quantization Mutual Information Maximization
$\mathrm{Zheng~Chen^{\star}1,~Lingwei~Zhu^{\star2},~Ziwei~Yang^{3},~and~Takashi~Matsubara}^{1}$
$\begin{array}{c}\\{^{1}\text{Osaka University, Japan}}\\{^{2}\text{University of Alberta, Canada}}\\{^{3}\text{Nara Institute of Science... | [
[
0.26708984375,
0.1458740234375,
0.73876953125,
0.1458740234375,
0.73876953125,
0.1634521484375,
0.26708984375,
0.1634521484375
],
[
0.2279052734375,
0.1685791015625,
0.77783203125,
0.1685791015625,
0.77783203125,
0.18310546875,
0.2279052734375,
0... | [
"Automated Cancer Subtyping via Vector",
"Quantization Mutual Information Maximization",
"Zheng Chen*1, Lingwei Zhu*2, Ziwei Yang3, and Takashi Matsubaral",
"1 Osaka University, Japan",
"2 University of Alberta, Canada",
"3 Nara Institute of Science and Technology",
"chen.zheng.bn1@gmail.com",
"Abstra... | {
"arxiv_id": "2206.10801",
"language": "zh",
"page_idx": 0
} | ||
2206.10801-zh-1 | 2206.10801-zh-1.jpg | 2 F. Author et al.
情况可以作为一个聚类问题[11]来解决，其中聚类给出了潜在亚型的提示。这种聚类方法主要依赖于数据的质量和合适的表示。现代分型方法通常利用分子转录组表达谱 (简称表达谱), 其由表征癌症特性的遗传和 microRNA (miRNA) 表达组成[26,21] 。但是，在充分利用表达式配置文件的功能方面存在一些困境： - High-dimensionality:the expression profiles are typically of>60,000dimen-
sions; even after typical preprocessing the dimension can still be >... | [
[
0.21728515625,
0.117431640625,
0.268798828125,
0.117431640625,
0.268798828125,
0.1287841796875,
0.21728515625,
0.1287841796875
],
[
0.26953125,
0.11614990234375,
0.37744140625,
0.11614990234375,
0.37744140625,
0.1300048828125,
0.26953125,
0.13000... | [
"2",
"F. Author et al.",
"情况可以作为一个聚类问题[11]来解决，其中聚类给出了潜在亚型的提示。这",
"种聚类方法主要依赖于数据的质量和合适的表示。现代分型方法通常利用分子",
"转录组表达谱（简称表达谱），其由表征癌症特性的遗传和microRNA（miRNA）",
"表达组成26.211。但是，在充分利用表达式配置文件的功能方面存在一些困境",
"-",
"High-dimensionality : the expression profiles are typically of > 60, 000 dimen",
"sions; even after typic... | {
"arxiv_id": "2206.10801",
"language": "zh",
"page_idx": 1
} | ||
2206.10801-zh-2 | 2206.10801-zh-2.jpg | ${\mathrm{Abbreviated~paper~title}}$ 3
(Effective) We compare the clustering results of VQ-RIM against existing ground truth labels (together with controversial labels from the entirety of labels) on different cancer datasets and find that VQ-RIM concords well with the ground truth, which verifes the correctness of VQ... | [
[
0.57666015625,
0.11614990234375,
0.7890625,
0.11614990234375,
0.7890625,
0.1300048828125,
0.57666015625,
0.1300048828125
],
[
0.2279052734375,
0.146484375,
0.78515625,
0.1483154296875,
0.78515625,
0.1634521484375,
0.2279052734375,
0.1610107421875... | [
"Abbreviated paper title3",
"-(Effective)We compare the clustering results of VQ-RIM against existing",
"ground truth labels (together with controversial labels from the entirety of",
"labels) on different cancer datasets and find that VQ-RIM concords well with",
"the ground truth, which verifies the correc... | {
"arxiv_id": "2206.10801",
"language": "zh",
"page_idx": 2
} | ||
2206.10801-zh-3 | 2206.10801-zh-3.jpg | 4 F. Author et al.
# 3 Method
# 3.1 Problem Setting
令$\mathcal{X}$为数据集$\mathcal{X}=\{x_1,\ldots,x_N\}$ ,其中$\boldsymbol{x}_i\in\mathbb{R}^d,1\leq i\leq N$是由癌症表达谱组成的$d$维载体。对于给定的$x$ ,我们的目标在于确定给定$x$的合适的癌症亚型$y\in \{ 1, 2, \ldots , K\}$ ,其中$K$不是预先固定的 ,需要自动确定。诸如$y=1,\ldots,K$ 之类的数值本身不具有任何医学解释，并且由于基础数据而简单地表示不同的表示。值得注意的是，虽然可... | [
[
0.223876953125,
0.11297607421875,
0.2344970703125,
0.12060546875,
0.223876953125,
0.1307373046875,
0.2132568359375,
0.12371826171875
],
[
0.27294921875,
0.11614990234375,
0.37744140625,
0.11614990234375,
0.37744140625,
0.1287841796875,
0.27294921875,... | [
"N",
"F. Author et al.",
"3",
"Method",
"3.1",
"Problem Setting",
"令为数据集={ai,.,a}，其中a,ERd,1≤i≤N是由癌症表达谱组",
"成的d维载体。对于给定的&，我们的目标在于确定给定&的合适的癌症亚型",
"yE{1,2,..,K}，其中K不是预先固定的，需要自动确定。诸如y=1,..,K",
"之类的数值本身不具有任何医学解释，并且由于基础数据而简单地表示不同的",
"表示。值得注意的是，虽然可以使用标签集，但它包含一小部分经过医学验证",
"的真实标签Vgt=（ygt}，以及大部分有争议的标签V=... | {
"arxiv_id": "2206.10801",
"language": "zh",
"page_idx": 3
} | ||
2206.10801-zh-4 | 2206.10801-zh-4.jpg | Abbreviated paper title 5
$\cdot$ Marginal entropy $\widehat{\mathcal{H}}\left(P(y)\right)$ aims to separate the subtypes as far as possi-
ble. Intuitively, it attempts to keep the subtypes uniform . Maximizing only $\widehat{\mathcal{H}}\left(P(y|\mathcal{X})\right)\stackrel{\text{(-1)}}{\operatorname*{\operatorname*... | [
[
0.57666015625,
0.11614990234375,
0.736328125,
0.11614990234375,
0.736328125,
0.1300048828125,
0.57666015625,
0.1300048828125
],
[
0.7294921875,
0.117431640625,
0.78759765625,
0.117431640625,
0.78759765625,
0.1287841796875,
0.7294921875,
0.1287841... | [
"Abbreviated paper title",
"5",
": Marginal entropy H(P(y) aims to separate the subtypes as far as possi-",
"ble. Intuitively, it attempts to keep the subtypes uniform . Maximizing only",
"H (P(ylx)) tends to produce degenerate solutions by removing subtypes",
"[6,19] , hence H (P(y)) serves as an effecti... | {
"arxiv_id": "2206.10801",
"language": "zh",
"page_idx": 4
} | ||
2106.01195-zh-0 | 2106.01195-zh-0.jpg | ## 识别文本附属物的比喻语言
Tuhin Chakrabarty$^*1,\textbf{Debanjan Ghosh}^*{3}$,Adam Poliak$^2,4$ and Smaranda Muresan$^{1,4}$
$^{1}$Department of Computer Science, Columbia University
$^{2}$Department of Computer Science, Barnard College
$^{3}$Educational Testing Service, $^{4}$Data Science Institute, Columbia University
$\{$ tu... | [
[
0.35693359375,
0.0831298828125,
0.64208984375,
0.0843505859375,
0.64208984375,
0.1015625,
0.35693359375,
0.1009521484375
],
[
0.1368408203125,
0.12469482421875,
0.873046875,
0.12469482421875,
0.873046875,
0.1395263671875,
0.1368408203125,
0.13952... | [
"识别文本附属物的比喻语言",
"Tuhin Chakrabarty*, Debanjan Ghosh * 3, Adam Poliak2,4 and Smaranda Muresan1,4",
"IDepartment ofComputer Science,Columbia University",
"2Department of Computer Science, Barnard College",
"3Educational Testing Service, *Data Science Institute, Columbia University",
"( tuhin.chakr, smara ) ... | {
"arxiv_id": "2106.01195",
"language": "zh",
"page_idx": 0
} | ||
2106.01195-zh-1 | 2106.01195-zh-1.jpg | <table>
<tbody>
<tr>
<td>$\overline{\text{Data}}$</td>
<td>$\gimel$</td>
<td>$\overline{\mathrm{Total}}$</td>
<td>$\overline{\mathrm{E}}$</td>
<td>$\overline{\mathrm{NE}}$</td>
</tr>
<tr>
<td>Simile</td>
<td>le</td>
<td>600</td>
<td>$\overline{300}$</td>
<td>$\overline{300}$</td>
... | [
[
0.20654296875,
0.0736083984375,
0.2376708984375,
0.0736083984375,
0.2376708984375,
0.08489990234375,
0.20654296875,
0.08489990234375
],
[
0.333251953125,
0.07598876953125,
0.366943359375,
0.07598876953125,
0.366943359375,
0.0843505859375,
0.333251953... | [
"Data",
"Total",
"NE",
"翻转来创建未附尾的示例，并将原始（Literal，",
"Simile",
"",
"600",
"300",
"300",
"613",
"307",
"306",
"Simile）对用作附尾。例如，在现有的上下文",
"Metaphon",
"SIGN2000",
"2.000",
"1867",
"假设对表达Entailment的情况下-“以这种方式",
"IronyMeaning",
"133",
"4,762",
"4.762",
"上台的普通公民就像绿色的月亮。“→”以这",
... | {
"arxiv_id": "2106.01195",
"language": "zh",
"page_idx": 1
} | ||
2106.01195-zh-3 | 2106.01195-zh-3.jpg | <table>
<tbody>
<tr>
<th> </th>
<th> </th>
<th> </th>
<th>Gold</th>
<th>Pred</th>
</tr>
<tr>
<td rowspan="3">Simile</td>
<td>一</td>
<td>Your guardian angel is just a little too much like a nerd at a comic convention . Your guardian angel is just a little too enthusiastic</td>
<td>$√$</... | [
[
0.78662109375,
0.0748291015625,
0.83447265625,
0.0748291015625,
0.83447265625,
0.08905029296875,
0.78662109375,
0.08905029296875
],
[
0.83447265625,
0.07598876953125,
0.8759765625,
0.07598876953125,
0.8759765625,
0.08905029296875,
0.83447265625,
... | [
"Gold",
"Pred",
"Yourguardiana",
"angelisjustalittletoomuchlikeanerdatacomicconvention",
"V",
"+",
"Your guardian angel is just a little too enthusiastic",
"Growing up, people always thought you were like a social pariah",
"Simile",
"x",
"V",
"Growing up,people always thought you were ordinary... | {
"arxiv_id": "2106.01195",
"language": "zh",
"page_idx": 3
} | ||
1809.01361-zh-0 | 1809.01361-zh-0.jpg | 一种用于多域图像转换和操作的统一特征解纠缠器
Alexander H. Liu$^{1}$ Yen-Cheng Liu$^{2}$ Yu-Ying Yeh$^{3}$ Yu-Chiang Frank Wang$^{1,4}$
$^{1}$National Taiwan University, Taiwan
$^{2}$Georgia Institute of Technology, USA$\:^3$University of California, San Diego, USA $^{4}$MOST Joint Research Center for AI Technology and All Vista Healthcare... | [
[
0.191162109375,
0.1307373046875,
0.8037109375,
0.1307373046875,
0.8037109375,
0.1458740234375,
0.191162109375,
0.1458740234375
],
[
0.1895751953125,
0.203857421875,
0.32177734375,
0.203857421875,
0.32177734375,
0.2171630859375,
0.1895751953125,
0... | [
"一种用于多域图像转换和操作的统一特征解纠缠器",
"Alexander H. Liul",
"Yen-Cheng Liu2",
"Yu-Ying Yeh3",
"",
"Yu-Chiang Frank Wangl,4",
"INational Taiwan University,Taiwan",
"2Georgia Institute of Technology, USA",
"3University of California, San Diego, USA",
"4MOsT JointResearch CenterforAITechnology andAll VistaHealthc... | {
"arxiv_id": "1809.01361",
"language": "zh",
"page_idx": 0
} | ||
1809.01361-zh-1 | 1809.01361-zh-1.jpg | Domain-Invariant
Latent Space
<FigureHere>
Pixel Space (Discrete Domains)
Pixel Space (Continuous Domains)
Figure 1: Illustration of multi-domain image translation and manipulation. With data from different $\tilde{\text{domains (e.g., }D_1:\text{sketch, }D_2:\text{photo, }D_3:\text{painting), the goal is to learn d... | [
[
0.41015625,
0.0953369140625,
0.53173828125,
0.0953369140625,
0.53173828125,
0.10919189453125,
0.41015625,
0.10919189453125
],
[
0.68310546875,
0.0953369140625,
0.7607421875,
0.09783935546875,
0.7607421875,
0.111083984375,
0.68310546875,
0.1091918... | [
"Domain-Invariant",
"Pixel Space",
"LatentSpace",
"Pixel Space",
"(Discrete Domains)",
"(Continuous Domains)",
"Underlyingfactors",
"",
"colorsaturation",
"X",
"sharpness",
"",
"contrast",
"Unified",
"",
"",
"Unified",
"Encoder",
"Generator",
"8",
"",
"X3",
"Manipulate",
... | {
"arxiv_id": "1809.01361",
"language": "zh",
"page_idx": 1
} | ||
1809.01361-zh-2 | 1809.01361-zh-2.jpg | 抗训练的方法，使模型对域移动不变。Tzeng 等人。[26]还试图通过引人域混淆损失来构建域不变分类器。通过推进对抗性学习策略，Bousmalis 等人。[2]选择学习正交表示，分别由共享和特定于域的编码器导出。Tzeng 等人。[27]通过对抗性训练在源域和目标域之间调整 CNN 特征提取器$\prime$分类器来解决 UDA 问题。然而，上述方法通常通过消除域偏差来解决域自适应问题。不能保证派生的表示会保留语义信息(例如。g。感兴趣的域或属性)。此外，由于 UDA 的目标是视觉分类，因此无法轻松实现图像翻译 (双域或多域)。正如我们在小节中强调的那样。1,我们的 UFDN 学习多域解纠缠表示，这使得多域图像到图像的翻译和操纵以及... | [
[
0.17724609375,
0.0947265625,
0.82177734375,
0.0947265625,
0.82177734375,
0.10479736328125,
0.17724609375,
0.10479736328125
],
[
0.176513671875,
0.106689453125,
0.82275390625,
0.106689453125,
0.82275390625,
0.12060546875,
0.176513671875,
0.1206054... | [
"抗训练的方法，使模型对域移动不变。Tzeng等人。[26]还试图通过引人域混淆损失来构建",
"域不变分类器。通过推进对抗性学习策略，Bousmalis等人。[2]选择学习正交表示，分别",
"由共享和特定于域的编码器导出。Tzeng等人。[271通过对抗性训练在源域和目标域之间",
"调整CNN特征提取器/分类器来解决UDA问题。然而，上述方法通常通过消除域偏差来解",
"决域自适应问题。不能保证派生的表示会保留语义信息（例如。g。、感兴趣的域或属性)。此",
"外，由于UDA的目标是视觉分类，因此无法轻松实现图像翻译（双域或多域）。正如我们在",
"小节中强调的那样。1，我们的UFDN学习多域解... | {
"arxiv_id": "1809.01361",
"language": "zh",
"page_idx": 2
} | ||
1809.01361-zh-3 | 1809.01361-zh-3.jpg | <FigureHere>
Figure 2: Overview of our Unified Feature Disentanglement Network (UFDN), consisting of an encoder $E$ ,a generator $G$ ,a discriminator in pixel space $D_x$ and a discriminator in feature space $D_{v}$ . Note that $x_c$ and $\hat{x}_c$ denote input and reconstruct images with domain vector $v_c$ ,respect... | [
[
0.275390625,
0.10101318359375,
0.340576171875,
0.10101318359375,
0.340576171875,
0.11236572265625,
0.275390625,
0.11236572265625
],
[
0.35693359375,
0.1307373046875,
0.414306640625,
0.1307373046875,
0.414306640625,
0.1383056640625,
0.35693359375,
... | [
"PixelSpace",
"Reconstruct",
"Real/Fake",
"x",
"X",
"Domain",
"Classifier",
"G",
"Adversarial",
"D",
"Domain",
"Classifier",
"FeatureSpace",
"Figure 2:Overview of our Unified Feature Disentanglement Network (UFDN), consisting of an",
"encoder E , a generator G , a discriminator in pixel ... | {
"arxiv_id": "1809.01361",
"language": "zh",
"page_idx": 3
} | ||
2208.10625-zh-0 | 2208.10625-zh-0.jpg | 1
<NEXTHT\_LINE
教育数据挖掘(EDM)应用数据挖掘，人工智能(AI)和机器学习(ML)来改善学术体验。近年来，许多教育机构广泛研究和部署了注入 AI 的技术[3,19] 。EDM 中最吸引人的重要任务之一是学生成绩预测。学生学习成果的早期估计可以帮助检测和通知学生在学业失败的风险。此外，它支持机构管理人员确定影响学生成绩的关键因素，并提供适当的干预措施以改善结果。成绩预测过程依赖于历史学术记录，并在标记数据上训练 ML 算法以预测学生的成绩。为此， 已经提出了各种数据集[11,26,37]和方法[16,24,41] 。随着人工智能系统的广泛使用和好处，公平性已成为设计此类系统的关键标准。非歧视性 ML 模型已成为教... | [
[
0.77685546875,
0.11932373046875,
0.7861328125,
0.11932373046875,
0.7861328125,
0.126953125,
0.77685546875,
0.126953125
],
[
0.2459716796875,
0.1483154296875,
0.3798828125,
0.147705078125,
0.3798828125,
0.1610107421875,
0.2459716796875,
0.16223144... | [
"1",
"<NEXTHTLINE",
"教育数据挖掘（EDM）应用数据控掘：人T智能（AI）和机器学习（ML）来",
"改善学术体验。近年来，许多教育机构广泛研究和部署了注入AI的技术[3,19]",
"EDM中最吸引人的重要任务之一是学生成绩预测。学生学习成果的早期估计",
"可以帮助检测和通知学生在学业失败的风险。此外，它支持机构管理人员确定",
"影响学生成绩的关键因素，并提供适当的干预措施以改善结果。成绩预测过程",
"依赖于历史学术记录，并在标记数据上训练ML算法以预测学生的成绩。为此",
"已经提出了各种数据集[11,26,37]和方法[16,24,41]。随着人工智能系统的广"... | {
"arxiv_id": "2208.10625",
"language": "zh",
"page_idx": 0
} | ||
2208.10625-zh-1 | 2208.10625-zh-1.jpg | 2
为分类引人了 20 多个公平概念[28,36] 。最著名的公平衡量标准之一是人口均等，即所谓的统计均等。它要求在受保护和不受保护的群体中进行正面预测的概率相等。然而，Dwork 等人。[13] 认为该指标无法确保个人公平。为了避免这种情况，Hardt 等人。[18] 提出的均衡赔率度量。它测量分类器是否同样好地预测所有属性值的标签。此外，还引人了许多其他流行的指标，并将其用于公平性 ML 研究，例如预测奇偶性，预测相等性[E[9] ,治疗相等性[E[8] 等。尽管有大量的公平措施，但没有适合所有情况的指标[28,36]
随着公平措施的发展，最近的研究试图在教育背景下评估公平[17,22,39] 。安德森等人。[6] 对现有学生... | [
[
0.2181396484375,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.2181396484375,
0.1275634765625
],
[
0.2459716796875,
0.149658203125,
0.78173828125,
0.149658203125,
0.78173828125,
0.160400390625,
0.24597167968... | [
"2",
"为分类引入了20多个公平概念[28.36]。最著名的公平衡量标准之一是人",
"口均等，即所谓的统计均等。它要求在受保护和不受保护的群体中进行正面预",
"测的概率相等。然而，Dwork等人。[13]认为该指标无法确保个人公平。为了",
"避免这种情况，Hardt等人。[18]提出的均衡赔率度量。它测量分类器是否同",
"样好地预测所有属性值的标签。此外，还引人了许多其他流行的指标，并将其用",
"于公平性ML研究，例如预测奇偶性，预测相等性E[9]，治疗相等性E[8]等。",
"尽管有大量的公平措施，但没有适合所有情况的指标[28,36]",
"随着公平措施的发展，最近的研究试图... | {
"arxiv_id": "2208.10625",
"language": "zh",
"page_idx": 1
} | ||
2208.10625-zh-2 | 2208.10625-zh-2.jpg | 3
${\textbf{Predicted class}}$
<FigureHere>
Fig.1: The confusion matrix with an example
### 2.1 Statistical parity
统计奇偶性(表示为 SP) 是众所周知的组公平性度量[13] ,由此，如果所研究的任何两个组之间的预测结果($\hat{Y}$)的差异(偏差),则任何分类器的输出都满足统计奇偶性(i。e。, $s$和$\overline{s})$达到了预定义的容差阈值$\epsilon:$
$$P(\hat Y|S=s)-P(\hat Y|S=\overline{s})\leq\epsilon.$$
(1)
我... | [
[
0.775390625,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1287841796875,
0.775390625,
0.1287841796875
],
[
0.50341796875,
0.1514892578125,
0.61767578125,
0.1514892578125,
0.61767578125,
0.1622314453125,
0.50341796875,
0.162231445... | [
"3",
"Predicted class",
"Positive +",
"Negative -",
"True Positive (TP)",
"False Negative (FN)",
"Positive +",
"TPprot + TPnon-prot",
"FNprot+FNnon-prot",
"10 (4:6)",
"",
"70(32:38)",
"False Positive (FP)",
"True Negative (TN)",
"Negative -",
"FPprot + FPnon-prot",
"TNprot + TNnon-pr... | {
"arxiv_id": "2208.10625",
"language": "zh",
"page_idx": 2
} | ||
2208.10625-zh-3 | 2208.10625-zh-3.jpg | 4
### 2.3 Equalized odds
预测器$\hat{Y}$满足均衡赔率(表示为 EOd)w。r。t。受保护的属性$S$ 和类标签$Y$,如果“$\hat{Y}$ 和$S$ 独立于$Y$”[18]。具体来说，男性和女性学生群体的预测真阳
性和假阳性概率应该相同。
$$P(\hat Y=+|S=s,Y=y)=P(\hat Y=+|S=\overline{s},Y=y),\quad y\in\{+,-\}.$$
由此，我们可以将均衡赔率测量为以下[21,27]:
$$EOd=\sum_{y\in\{+,-\}}|P(\hat Y=+|S=s,Y=y)-P(\hat Y=+|S=\overline{s},Y=y)|.$$... | [
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1287841796875,
0.21728515625,
0.1287841796875
],
[
0.21728515625,
0.147705078125,
0.3896484375,
0.147705078125,
0.3896484375,
0.1610107421875,
0.21728515625,
0... | [
"4",
"2.3Equalized odds",
"预测器Y满足均衡赔率（表示为EOd)W。r。t。受保护的属性S和类标签Y",
"如果“Y和S独立于Y”[18]。具体来说，男性和女性学生群体的预测真阳",
"性和假阳性概率应该相同。",
"P(Y=+IS =s,Y=y)=P(Y =+IS=s,Y=y),ye(+,-):",
"(5",
"由此，我们可以将均衡赔率测量为以下[21,27]：",
"EOd=",
" IP(Y = +IS= 8, Y =y) -P(Y= +IS =3,Y =y)]. (6)",
"ye(+,-}",
"值范围：EOdE[0,2]；0表示无判别，2表示... | {
"arxiv_id": "2208.10625",
"language": "zh",
"page_idx": 3
} | ||
2208.10625-zh-4 | 2208.10625-zh-4.jpg | 5
2.6 Treatment equality
如果保护组和非保护组的假阴性和假阳性的比率相同，则满足治疗平等[8](表
示为 TE)。
$$\begin{aligned}\frac{FN_{prot.}}{FP_{prot.}}=\frac{FN_{non-prot.}}{FP_{non-prot.}}.\end{aligned}$$
(11)
在我们的论文中，我们通过方程 11 中描述的两个比率之间的差来报告治疗平等。如果$FP_prot.$或$FP_non-prot.\text{ 为零 }^2$,则度量变为无界。在我们的例子中， $TE= - 0. 2$ , 因 为 FN 和 FP 的 比 率 是 1 和 1。女 性... | [
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1287841796875,
0.77392578125,
0.1287841796875
],
[
0.2181396484375,
0.1483154296875,
0.26318359375,
0.1483154296875,
0.26318359375,
0.160400390625,
0.2181396484375,
0.16... | [
"5",
"2.6",
"Treatment equality",
"如果保护组和非保护组的假阴性和假阳性的比率相同，则满足治疗平等[8（表",
"示为TE)。",
"FNprot.-FNnon-prot.",
"(11)",
"FPprot.",
"FPnon-prot",
"在我们的论文中，我们通过方程11中描述的两个比率之间的差来报告治疗平",
"等。如果FPprot.或FPnon-prot.为零？，则度量变为无界。在我们的例子中，",
"TE=-0.2，因为FN和FP的比率是1和1。女性和男性分别为2",
"2.7Absolute Between-ROCArea",
... | {
"arxiv_id": "2208.10625",
"language": "zh",
"page_idx": 4
} | ||
2210.09803-zh-1 | 2210.09803-zh-1.jpg | 用于情感分析的情感感知单词和句子级别预训练
Shuai Fan$^1,\textbf{ Chen Lin}^1;$ Haonan Li$^2\dagger\textbf{ Zhenghao Lin}^1\dagger,\textbf{ Jinsong Su}^1\textbf{ Hang Zhang}^3$,
Yeyun Gong$^4$,Jian Guo$^3$,Nan Duan$^4$
$^{1}$ School of Informatics, Xiamen University, China
$^{2}$The University of Melbourne, Australia
$^{3}$ IDEA Research... | [
[
0.2493896484375,
0.09442138671875,
0.75,
0.09442138671875,
0.75,
0.1092529296875,
0.2493896484375,
0.1092529296875
],
[
0.145263671875,
0.127685546875,
0.8623046875,
0.127685546875,
0.8623046875,
0.142578125,
0.145263671875,
0.142578125
],
[
... | [
"用于情感分析的情感感知单词和句子级别预训练",
"Shuai Fan', Chen Lin', Haonan Li2, Zhenghao Lin', Jinsong Su' Hang Zhang,",
"Yeyun Gong', Jian Guo3, Nan Duan",
"1 School of Informatics, Xiamen University, China",
"2 The University of Melbourne, Australia",
"3 IDEA Research, China",
"4 Microsoft Research Asia",
"Abstract",
... | {
"arxiv_id": "2210.09803",
"language": "zh",
"page_idx": 1
} | ||
2210.09803-zh-2 | 2210.09803-zh-2.jpg | 通过屏蔽的情感词生成和检测来加强词级预训练，样本效率更高，有利于各种情感分类任务；2)。SentiWSP 将单词级预训练与句子级预 对属于下游任务相同域的语料库进行后训练，
训练相结合，这在以前的研究中没有得到充分考虑。SentiWSP在预训练中采用对比学习，将句子与批量和跨批次的硬阴性进行逐步对比， 从而使模型能够对句子的详细情感信息进行编码；3)。我们对句子级和方面级情感分类任务进行了广泛的实验，并表明 SentiWSP 在多个基准数据集上实现了最新的性能。
## 2 $\textbf{Related Work}$
Pre-training and Representation Learning 预训练模型在各种 NLP 任... | [
[
0.12005615234375,
0.0885009765625,
0.486083984375,
0.0885009765625,
0.486083984375,
0.09857177734375,
0.12005615234375,
0.09857177734375
],
[
0.5224609375,
0.0872802734375,
0.87890625,
0.0872802734375,
0.87890625,
0.1009521484375,
0.5224609375,
0... | [
"通过屏蔽的情感词生成和检测来加强词级预",
"Pre-trained Models for Sentiment Analysis",
"训练，样本效率更高，有利于各种情感分类任",
"在情感分析领域，BERT-PTE(Xuetal.，2019",
"务；2)。SentiWSP将单词级预训练与句子级预",
"对属于下游任务相同域的语料库进行后训练，",
"训练相结合，这在以前的研究中没有得到充分",
"以受益于方面级情感分类。SKEP（Tianetal.,",
"考虑。SentiWSP在预训练中采用对比学习，将",
"2020）构建三个情感知识预测目标，以便为多",
"句子与批量和跨批... | {
"arxiv_id": "2210.09803",
"language": "zh",
"page_idx": 2
} | ||
2210.09803-zh-3 | 2210.09803-zh-3.jpg | <FigureHere>
Figure 1: Framework overview of SentiWSP.
Discriminator 对于损坏的句子，鉴别器$D$ ,i。e。,一个更大尺寸的 Transformer,将损坏的句子编码为$h_{\boldsymbol{D}}(s^\mathrm{rep})$,并使用 sigmoid 输出层预测每个单词$w_t$是来自数据还是来自生成器：
$$D(s^\text{rep},t)=\sigma\left(e_t^Th_{\boldsymbol{D}}(s^\text{rep})_t\right)$$
(2)
我们共同训练生成器和鉴别器。通过最大似然估计训练生成器$G$ ,并通过交... | [
[
0.193115234375,
0.0885009765625,
0.2091064453125,
0.0885009765625,
0.2091064453125,
0.0968017578125,
0.193115234375,
0.0968017578125
],
[
0.7109375,
0.09027099609375,
0.74072265625,
0.09027099609375,
0.74072265625,
0.097412109375,
0.7109375,
0.09... | [
"0",
"Index",
"Index",
"O:original",
"&Search",
"A",
"A",
"4",
"4",
"A",
"&Search",
"R:replaced",
"Inferencing",
"Inferencing",
"Inferencer",
"Discriminator",
"4",
"4",
"Asmanstylish",
"exceptionally",
"charming",
"book",
"ANN",
"D/k-1",
"Drk",
"Drk~2",
"Negatives... | {
"arxiv_id": "2210.09803",
"language": "zh",
"page_idx": 3
} | ||
2210.09803-zh-4 | 2210.09803-zh-4.jpg | <table>
<tbody>
<tr>
<th rowspan="3">$\mathbf{Model}$</th>
<th colspan="4">Sentence-level</th>
<th> </th>
<th colspan="4">Aspect- level</th>
</tr>
<tr>
<th rowspan="2">IMDB Acc</th>
<th rowspan="2">$SST-5$ Acc</th>
<th rowspan="2">$Yelp-2$ Acc</th>
<th rowspan="2">$Yelp-5$ Acc</th>
<th... | [
[
0.450927734375,
0.0872802734375,
0.55517578125,
0.0872802734375,
0.55517578125,
0.1015625,
0.450927734375,
0.1015625
],
[
0.7255859375,
0.08905029296875,
0.81005859375,
0.08905029296875,
0.81005859375,
0.1009521484375,
0.7255859375,
0.10095214843... | [
"Sentence-level",
"Aspect-level",
"IMDB",
"Yelp-2",
"Yelp-5",
"Restaurant14",
"Laptop14",
"SST-5",
"Model",
"MR",
"Acc",
"Acc",
"Acc",
"Acc",
"Acc",
"Acc",
"MF1",
"Acc",
"MF1",
"70.16",
"87.52",
"76.06",
"73.11",
"78.53",
"BERT (Devlin et al., 2019)",
"93.87",
"53... | {
"arxiv_id": "2210.09803",
"language": "zh",
"page_idx": 4
} | ||
1612.01887-zh-0 | 1612.01887-zh-0.jpg | Knowing When to Look: Adaptive Attention via
A Visual Sentinel for Image Captioning
Jiasen Lu $^2*\dagger$ Caiming Xiong $^1\dagger$ Devi Parikh$^3$, Richard Socher$^1$
$^{1}\text{Salesforce Research,}^{2}\text{Virginia Tech,}^{3}$Georgia Institute of Technology
jiasenlu@vt.edu, parikh@gatech.edu, $\{$ cxiong, rsocher... | [
[
0.24267578125,
0.1351318359375,
0.7294921875,
0.1351318359375,
0.7294921875,
0.149658203125,
0.24267578125,
0.149658203125
],
[
0.285888671875,
0.1546630859375,
0.685546875,
0.1578369140625,
0.685546875,
0.175537109375,
0.285888671875,
0.17236328... | [
"Knowing When to Look: Adaptive Attention via",
"A Visual Sentinel for Image Captioning",
"Jiasen Lu 2 *t Caiming Xiong 1 t Devi Parikh 3 , Richard Socher 1",
"1 Salesforce Research, 2 Virginia Tech, 3 Georgia Institute of Technology",
"jiasenlu@vt.edu, parikh@gatech.edu, [ cxiong, rsocher @salesforce.com"... | {
"arxiv_id": "1612.01887",
"language": "zh",
"page_idx": 0
} | ||
1612.01887-zh-1 | 1612.01887-zh-1.jpg | 总的来说，本文的主要贡献是：
$\bullet$ We introduce an adaptive encoder-decoder frame-
work that automatically decides when to look at the image and when to rely on the language model to generate the next word.
$\bullet$ We frst propose a new spatial attention model,
and then build on it to design our novel adaptive
attention model ... | [
[
0.101318359375,
0.0947265625,
0.31787109375,
0.0947265625,
0.31787109375,
0.10479736328125,
0.101318359375,
0.10479736328125
],
[
0.51708984375,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.51708984375,
0.1060791015625... | [
"总的来说，本文的主要责献是：",
". First, in the vanilla framework, Ct is only depend-",
" We introduce an adaptive encoder-decoder frame-",
"ent on the encoder, a Convolutional Neural Net-",
"work that automatically decides when to look at",
"work (CNN). The input image I is fed into the",
"the image and when to rel... | {
"arxiv_id": "1612.01887",
"language": "zh",
"page_idx": 1
} | ||
1612.01887-zh-3 | 1612.01887-zh-3.jpg | 。我们使用单层神经网络将视觉哨兵向量$s_t$和 LSTM
输出向量$h_t$转换为具有$d$维度的新向量
培训详情。在我们的实验中，我们使用隐藏大小为512的单层 LSTM。我们使用 Adam 优化器，语言模型的基本学习率为 5e-4, CNN 的基本学习率为 1e-5。动量和重量衰减为 0。8 和0。999 分别。我们在 20 个时代之后对 CNN 网络进行微调。如果验证 CIDEr [26] 分数在过去 6 个时期内没有改善，我们将批次大小设置为 80, 并在早期停止的情况下训练多达 50 个时期。我们的模型可以在单个 TitanXGPU 上在 30 小时内进行训练。在对 COCO 和 Flickr30k 数据集的标题进行采样... | [
[
0.09722900390625,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.09722900390625,
0.10479736328125
],
[
0.50244140625,
0.09344482421875,
0.5947265625,
0.09344482421875,
0.5947265625,
0.10479736328125,
0.50244140625,
... | [
"我们使用单层神经网络将视觉哨兵向量St和LSTM",
"5.Results",
"输出向量ht转换为具有d维度的新向量",
"",
"5.1. Experiment Settings",
"培训详情。在我们的实验中，我们使用隐藏小为",
"512的单层LSTM。我们使用Adam优化器，语言模",
"我们用两个数据集进行实验：Flickr30k[35]和",
"COCO [16]",
"型的基本学习率为5e-4，CNN的基本学习率为1e-5。",
"Flickr30k包含从Flickr收集的31,783个图像。这",
"动量和重量衰减为0。8和0。999分别。我们在20个",
... | {
"arxiv_id": "1612.01887",
"language": "zh",
"page_idx": 3
} | ||
1612.01887-zh-4 | 1612.01887-zh-4.jpg | <table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th colspan="6">Flickr30k</th>
<th colspan="6">MS- COCO</th>
</tr>
<tr>
<th>$B-1$</th>
<th>$B-2$</th>
<th>$B-3$</th>
<th>$B-4$</th>
<th>METEOR</th>
<th>CIDEr</th>
<th>$B-1$</th>
<th>$B-2$</th>
<th>$B-3$</th>
<th>$B-4$</th>
... | [
[
0.36279296875,
0.0947265625,
0.4248046875,
0.0947265625,
0.4248046875,
0.10919189453125,
0.36279296875,
0.10919189453125
],
[
0.705078125,
0.0953369140625,
0.77392578125,
0.0953369140625,
0.77392578125,
0.106689453125,
0.705078125,
0.106689453125... | [
"Flickr30k",
"MS-COCO",
"Method",
"B-1",
"B-2",
"B-3",
"B-4",
"METEOR",
"CIDEr",
"B-1",
"B-2",
"B-3",
"B-4",
"METEOR",
"CIDEr",
"0.573",
"0.240",
"0.450",
"0.321",
"0.157",
"0.369",
"0.153",
"0.247",
"0.625",
"0.230",
"0.195",
"0.660",
"Hard-Attention [30]",
"... | {
"arxiv_id": "1612.01887",
"language": "zh",
"page_idx": 4
} | ||
2104.02646-zh-0 | 2104.02646-zh-0.jpg | Published as a conference paper at ICLR 2021
$\nabla Sim:$ 用于系统识别和可视化运动控制的可微分仿
真
https://gradsim.github.io
Krishna Murthy Jatavallabhula* $^{1,3.4}$, Miles Macklin$^{*2}$,Florian Golemo$^{1,3}$,Vikram Voleti$^{3,4}$, Linda Petrini$^{3}$,Martin Weiss$^{3,4}$,Breandan Considine$^{3,5},\text{Jérôme Parent-Lévesque}^{3... | [
[
0.1748046875,
0.034088134765625,
0.476318359375,
0.034088134765625,
0.476318359375,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.176513671875,
0.10418701171875,
0.8203125,
0.10479736328125,
0.8203125,
0.12249755859375,
0.176513671875,
0.1218... | [
"Published as a conference paper at ICLR 2021",
"VSim：用于系统识别和可视化运动控制的可微分仿",
"真",
"https://gradsim.github.io",
"Krishna Murthy Jatavallabhula* 1,3.4, Miles Macklin*2, Florian Golemo'3, Vikram Voleti3.4, Linda",
"Petrini3, Martin Weiss3.4, Breandan Considine3.5, Jerome Parent-Lévesq",
"le3.5,KevinXie2.6.7... | {
"arxiv_id": "2104.02646",
"language": "zh",
"page_idx": 0
} | ||
2104.02646-zh-1 | 2104.02646-zh-1.jpg | Published as a conference paper at ICLR 2021
<FigureHere>
Figure 2: $\nabla Sim:$ Given video observations of an evolving physical system (e), we randomly initialize scene object properties (a) and evolve them over time using a differentiable physics engine (b), which generates states . Our renderer(c) processes stat... | [
[
0.1748046875,
0.034088134765625,
0.47802734375,
0.034088134765625,
0.47802734375,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.2034912109375,
0.07574462890625,
0.250732421875,
0.07574462890625,
0.250732421875,
0.08587646484375,
0.2034912109375,
... | [
"Published as a conference paper at ICLR 2021",
"→Rollout",
"-Gradient",
"Camera Colors LightsK Textures",
"MassVertices",
"Friction",
"Forces",
"O",
"States",
"Loss",
"",
"C",
"",
"Elasticity",
"(d) Predicted",
"(e) Ground",
"(a) Object",
"(b)Physics Engine",
"(c) Renderer",
"... | {
"arxiv_id": "2104.02646",
"language": "zh",
"page_idx": 1
} | ||
2104.02646-zh-2 | 2104.02646-zh-2.jpg | $\underline{\text{Published as a conference paper at ICLR 2021}}$
$\nabla Sim$包含两个主要组件：可微分物理引擎，用于计算每个时刻场景的物理状态；可微分渲染器，用于将场景渲染为 2D 图像。与现有的可微分物理$\boxed{\mathrm{E}}$(Toussaint et al., 2018; de Avila Belbute-Peres et al., 2018; Song & Boularias, 2020b;a; Degrave et al., 2016; Wu et al., 2017a; Research, 2020 (accessed May... | [
[
0.1748046875,
0.034088134765625,
0.476318359375,
0.034088134765625,
0.476318359375,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.1748046875,
0.10418701171875,
0.82177734375,
0.10479736328125,
0.82177734375,
0.11932373046875,
0.1748046875... | [
"Published as a conference paper at ICLR 2021",
"VSim包含两个主要组件：可微分物理引擎，用于计算每个时刻场景的物理状态；可微分",
"谊染器，用于将场景染为2D图像。与现有的可微分物理E(Toussaintetal.,2018;deAvila",
"Belbute-Peres et al., 2018; Song & Boularias, 2020b;a; Degrave et al., 2016; Wu et al.,2017a",
"Research,2020（accessedMay15,2020;Huetal.,2020;Qiaoetal.,2020)... | {
"arxiv_id": "2104.02646",
"language": "zh",
"page_idx": 2
} | ||
2104.02646-zh-3 | 2104.02646-zh-3.jpg | Published as a conference paper at ICLR 2021
<table>
<tbody>
<tr>
<th rowspan="2">Approach</th>
<th rowspan="2">Mean abs.</th>
<th rowspan="2">Abs. Rel</th>
<th rowspan="2">Approach</th>
<th colspan="2">elasticity</th>
<th colspan="2">friction</th>
</tr>
<tr>
<th>$\overline{k_d}$</th>
<t... | [
[
0.1748046875,
0.034088134765625,
0.47802734375,
0.034088134765625,
0.47802734375,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.685546875,
0.10101318359375,
0.72802734375,
0.1029052734375,
0.72607421875,
0.1142578125,
0.6845703125,
0.11236572... | [
"Published as a conference paper at ICLR 2021",
"elasticity",
"friction",
"Approach",
"Abs, Rel.",
"mass",
"Approach",
"m",
"",
"kf",
"T",
"Average",
"1.7713",
"4.11570.4463",
"ke",
"r",
"3.71452.3410",
"0.1031",
"Average",
"0.2022",
"10.0007",
"2.5454",
"5.02410.5558",
... | {
"arxiv_id": "2104.02646",
"language": "zh",
"page_idx": 3
} | ||
2104.02646-zh-4 | 2104.02646-zh-4.jpg | $\underline{\text{Published as a conference paper at ICLR 2021}}$
<table>
<tbody>
<tr>
<th> </th>
<th colspan="3">Deformable solid FEM</th>
<th rowspan="3">Thin-shell (cloth) Per-particle velocity $v$</th>
</tr>
<tr>
<th> </th>
<th rowspan="2">Per- particle mass $m$</th>
<th>Material</th>
<... | [
[
0.1732177734375,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.435546875,
0.10101318359375,
0.564453125,
0.10101318359375,
0.564453125,
0.1142578125,
0.435546875,
0.114257... | [
"Published as a conference paper at ICLR 2021",
"Deformable solid FEM",
"Thin-shell (cloth)",
"Per-particlemass",
"Material properties",
"Per-particle velocity",
"m",
"H",
"X",
"U",
"RelMAE",
"Rel. MAE",
"Rel.MAE",
"Approach",
"Rel. MAE",
"DiffPhysics (3D Sup.)",
"0.032",
"0.0025",... | {
"arxiv_id": "2104.02646",
"language": "zh",
"page_idx": 4
} | ||
2208.07004-zh-1 | 2208.07004-zh-1.jpg | 人工智能促进全球气候合作：在 RICE-N 中对全球气候谈判、协议和长期合作进
# 行建模
Tianyu Zhang$^1,2,\textbf{Andrew Williams}^{1,2},\textbf{Soham Phade}^6,\textbf{Sunil Srinivasa}^{6*},\textbf{Yang Zhang}^2$,
$\textbf{Prateek Gupta}^{2,3,4},\textbf{Yoshua Bengio}^{1,2,5},\textbf{Stephan Zheng}^{6}$
$^{1}$Université de Montréal,$^2$Mila,$^3$University of... | [
[
0.09478759765625,
0.1256103515625,
0.90869140625,
0.1256103515625,
0.90869140625,
0.14013671875,
0.09478759765625,
0.14013671875
],
[
0.465576171875,
0.1458740234375,
0.53515625,
0.1458740234375,
0.53515625,
0.1622314453125,
0.465576171875,
0.162... | [
"人工智能促进全球气候合作：在RICE-N中对全球气候谈判、协议和长期合作进",
"行建模",
"Tianyu Zhangl.2, Andrew Williams!.2, Soham Phade, Sunil Srinivasa *, Yang Zhang?,",
"Prateek Gupta2.3.4, Yoshua Bengiol,2.5, Stephan Zheng",
"1 Université de Montréal, 2 Mila, 3 University of Oxford,",
"4 The Alan Turing Institute, 5 CIFAR, 6 Salesforce Res... | {
"arxiv_id": "2208.07004",
"language": "zh",
"page_idx": 1
} | ||
2208.07004-zh-2 | 2208.07004-zh-2.jpg | Contributions. 在本文中，我们提出了一个概念和实践框架来克服这些局限性，并对理性代理行为及其对气候变化的影响进行战略分析。我们还介绍了人工智能促进全球气候合作，这是一项社区倡议，旨在促进该领域的跨学科研究和现实世界的影响。具体来说，我们提供以下内容：
· we argue that machine learning (ML) offers an attractive
framework to analyze the strategic aspects of real world
climate negotiations and agreements;
· we argue this research requires... | [
[
0.0906982421875,
0.06817626953125,
0.47802734375,
0.0694580078125,
0.47802734375,
0.08331298828125,
0.0906982421875,
0.08209228515625
],
[
0.52392578125,
0.07135009765625,
0.91015625,
0.07135009765625,
0.91015625,
0.08209228515625,
0.52392578125,
... | [
"Contributions．，在本文中，我们提出了一个概念和实",
"其他领域的框架解决了后一个缺点，并受到基于代理的",
"践框架来克服这些局限性，并对理性代理行为及其对气",
"建模EBonabeau2002）作为自下而上的建模框架的启",
"候变化的影响进行战略分析。我们还介绍了人工智能促",
"发，并在MARL中取得了进步，以找到性能良好的策",
"进全球气候合作，这是一项社区倡议，旨在促进该领域",
"略(Zhengetal.2022）并训练人类级战略代理E(Silver",
"的跨学科研究和现实世界的影响。具体来说，我们提供",
"etal.2016:Vinyals etal... | {
"arxiv_id": "2208.07004",
"language": "zh",
"page_idx": 2
} | ||
2208.07004-zh-3 | 2208.07004-zh-3.jpg | Le Gléau et al. 2022)。因此，MARL 是一个有吸引力的框架来分析气候结果，同时考虑到战略行为。然而，以前的工作在很大程度上考虑了高度程式化的环境，尚未应用于丰富的校准气候经济模拟；我们的工作填补了这一空白
3 $\textbf{The RICE- N Integrated Assessment}$
# Model
我们介绍 RICE-N ,这是一个 IAM, 它进一步增强了RICE 的谈判协议框架，并且还包括 Lessmann, Marschinski, and Edenhofer (2009) 之后的国际贸易和关税)。因此，RICE-N 与现实世界共享气候，经济和社会特征
在 RICE-N 中，有$n$区... | [
[
0.08905029296875,
0.0706787109375,
0.473876953125,
0.0706787109375,
0.473876953125,
0.080810546875,
0.08905029296875,
0.080810546875
],
[
0.5390625,
0.07135009765625,
0.91015625,
0.07135009765625,
0.91015625,
0.08209228515625,
0.5390625,
0.082092... | [
"LeGleauetal2022）。因此MARL是一个有吸引l力的",
"活动组件封装了这些动态。在每个步骤中，它都执行",
"框架来分析气候结果，同时考虑到战略行为。然而，以",
"以下操作：每个地区的总产值是根据该地区的状况，特",
"前的工作在很大程度上考虑了高度程式化的环境，尚未",
"别是其资本投资，劳动力（或人口）和技术因素来计算",
"应用于丰富的校准气候经济模拟；我们的工作填补了这",
"的。净经济产出是由于全球气温上升造成的气候损害而",
"+空白",
"减少的总产出，以及该地区为缓解气候变化所付出的努",
"力的成本。。该地区消费的国内商品等于资本投资和出",
... | {
"arxiv_id": "2208.07004",
"language": "zh",
"page_idx": 3
} | ||
2109.10257-zh-1 | 2109.10257-zh-1.jpg | 个设计组件组合在一个作品中，克服了先前作品的缺点。首先，为了使用骨架的空间配置信息，我们将问题建模为端到端的时空图。我们依靠邻接矩阵来编码骨架关节之间的关系。而不是固定的邻接矩阵，我们让我们的模型学习它并在我们的消融研究中分析它。其次， 我们通过将视觉信号融合到我们的模型中来利用视觉信号。因此，我们利用上下文信息来增强我们的结果。我们发现视觉信号在某些情况下增强了短期预测。最后，为了避免长期的分歧，我们使用完整的 CNN 方法端到端，没有任何反复出现的行为。与以前的作品相比，这导致了更好的长期预测。我们引人了一种新的度量方法，该度量方法可以测量预测范围内的差异，以定量判断此预测稳定性标准
这项工作的组织方式如下，我们从相关的 2D ... | [
[
0.08416748046875,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.08416748046875,
0.10479736328125
],
[
0.50341796875,
0.09344482421875,
0.890625,
0.09344482421875,
0.890625,
0.106689453125,
0.50341796875,
0.10668945... | [
"个设计组件组合在一个作品中，克服了先前作品的缺",
"导致姿态和轨迹预测领域的飞跃[25，16,37，5，371，因",
"点。首先，为了使用骨架的空间配置信息，我们将问题",
"为图CNN可以利用代理的空间配置及其相应的内部特",
"建模为端到端的时空图。我们依靠邻接矩阵来编码骨",
"征。通过考虑骨骼E40，171的时空图公式：成功地对",
"架关节之间的关系。而不是固定的邻接矩阵，我们让我",
"人类骨骼的动作进行了分类。。[21]的最新工作提出了",
"们的模型学习它并在我们的消融研究中分析它。其次",
"",
"-种多尺度深度图网络，用于从3D历史预测3D人体",
"我... | {
"arxiv_id": "2109.10257",
"language": "zh",
"page_idx": 1
} | ||
2109.10257-zh-3 | 2109.10257-zh-3.jpg | # 4. Experiments
在本节中，我们将描述训练中使用的数据集，训练设置和评估指标。然后我们与几个先前的模型进行比较，然后以定量和定性的方式对我们的模型进行消融研究。
## 4.1. Datasets description
GTA-IM: GTA 室内运动数据集$\mathbb{E}$[6] 强调人与场景的交互。该数据集的动机是解决真实的人机交互数据集存在的问题，即嘈杂的 3D 人体姿势注释和有限的远程人体运动。通过控制角色，摄像机和物理系统，从 3D 视频游戏 GrandTheftAutoV 中收集了运动和交互的合成数据。数据集包含在 10 个不同的大型室内场景中表演的 50 个人类角色。数据集包括具有 1920... | [
[
0.08251953125,
0.09344482421875,
0.468017578125,
0.09344482421875,
0.468017578125,
0.106689453125,
0.08251953125,
0.106689453125
],
[
0.50341796875,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.106689453125,
0.50341796875,
0.1066894531... | [
"例如，骨架可能正在设置，但颈部和身体之间的距离很",
"总共记录了60个场景。每个视频为30FPS，具有摄像",
"尴尬。这些异常骨架的例子可以在图4中看到。因此，",
"机参数，校准参数和人体片段。人体姿势的3D骨架点",
"我们引入了一个损失函数，迫使预测的骨架更加一致。",
"（25个关节）由Kinect-One传感器捕获，2D关键点（25",
"我们将这种损失称为骨骼一致性损失（SCL），它在预测",
"个关节）由OpenPose捕获。在我们的实验中，我们将",
"的3D姿势中强制正确的骨骼长度和关节之间的角度。",
"PROX数据集分为52个训练序列和8个序列，用于按",
... | {
"arxiv_id": "2109.10257",
"language": "zh",
"page_idx": 3
} | ||
2109.10257-zh-4 | 2109.10257-zh-4.jpg | <FigureHere>
Figure 4: Different cases of deformation of the predicted 3D skeletons. Arrow indicates the time direction. The first two
poses (a) and (b) show awkward joint angles between each joint. The case (c) one the right shows failure in both predicted
angle and bone length.
倍。每 200 个时代 2 个。我们使用 128 的批量大小和1 秒的... | [
[
0.6337890625,
0.0921630859375,
0.7099609375,
0.0921630859375,
0.7099609375,
0.1029052734375,
0.6337890625,
0.1029052734375
],
[
0.63232421875,
0.11236572265625,
0.6953125,
0.11236572265625,
0.6953125,
0.12371826171875,
0.63232421875,
0.1237182617... | [
"groundtruth",
"prediction",
"alwalkino",
"(c) lying",
"/ (b) sitting",
"Figure 4: Different cases of deformation of the predicted 3D skeletons. Arrow indicates the time direction. The first two",
"poses (a) and (b) show awkward joint angles between each joint. The case (c) one the right shows failure i... | {
"arxiv_id": "2109.10257",
"language": "zh",
"page_idx": 4
} | ||
2103.16155-zh-1 | 2103.16155-zh-1.jpg | 通过
学习动作和上下文的显式子空间进行弱监督的时间动作本地化
Ziyi Liu$^1$, Le Wang$^1*$, Wei Tang$^2$,Junsong Yuan$^3$,Nanning Zheng$^1$,Gang Hua$^4\:^1$Institute of
Artificial Intelligence and Robotics, Xi`an Jiaotong University
$^{2}$University of Illinois at Chicago $^3\text{The State University'of New York at Buffalo}^{4}$Wormpex AI Research
$... | [
[
0.48193359375,
0.125,
0.52783203125,
0.125,
0.52783203125,
0.14013671875,
0.48193359375,
0.14013671875
],
[
0.1895751953125,
0.1427001953125,
0.82177734375,
0.143310546875,
0.82177734375,
0.1610107421875,
0.1895751953125,
0.160400390625
],
[
... | [
"通过",
"学习动作和上下文的显式子空间进行弱监督的时间动作本地化",
"Ziyi Liul, Le Wangl*, Wei Tang?, Junsong Yuan², Nanning Zheng', Gang Hua 1Institute of",
"Artificial Intelligence and Robotics, Xi'an Jiaotong University",
"2University of Illinois at Chicago 3The State University of New York at Buffalo 4Wormpex AI Research",
"liuziyi... | {
"arxiv_id": "2103.16155",
"language": "zh",
"page_idx": 1
} | ||
2103.16155-zh-2 | 2103.16155-zh-2.jpg | <FigureHere>
Figure 1: Illustration of a common WS-TAL pipeline with action-context confusion (left) and the main idea of our method (right). The original feature space jointly represents action and context visual elements, resulting in confusion of actions (red dots) and context (black dots). The proposed method lear... | [
[
0.529296875,
0.07135009765625,
0.67236328125,
0.07135009765625,
0.67236328125,
0.08209228515625,
0.529296875,
0.08209228515625
],
[
0.7861328125,
0.06756591796875,
0.890625,
0.06756591796875,
0.890625,
0.07763671875,
0.7861328125,
0.07763671875
... | [
"OriginalFeatureSpace",
"ActionSubspace",
"Input Untrimmed Video with T Snippets",
"Jointly",
"IRepresenting",
"Containing Action Baseball Pitch",
"[Action and",
"Snippet with Action",
"Feature Embedding",
"HContext",
" Snippet with Only",
"",
"b",
"",
"Context",
"",
"...",
"",
"... | {
"arxiv_id": "2103.16155",
"language": "zh",
"page_idx": 2
} | ||
2206.14467-zh-2 | 2206.14467-zh-2.jpg | <FigureHere>
Figure 1: Overview of our proposed Test-time Adaptation from Shape Dictionary (TASD). During training, we establish dictionary learning to extract the explicit shape priors from single domain data and integrate these prior information into the network by introducing a regression branch to produce shape co... | [
[
0.09234619140625,
0.0694580078125,
0.25244140625,
0.0706787109375,
0.25244140625,
0.08209228515625,
0.09234619140625,
0.080810546875
],
[
0.438720703125,
0.0706787109375,
0.69677734375,
0.0706787109375,
0.69677734375,
0.080810546875,
0.438720703125,
... | [
"Integration of Shape Dictionary",
"Test-time Update via Dual-consistency Regularizatio",
"Encode",
"Decode",
"",
"Noise!",
"",
"Regression branch",
"admai",
"Regression branch",
"Lseg",
"",
"",
"Shape",
"Ceon",
"coefficient",
"Shape coefficiemt conststenc",
"Ccons",
"prediction"... | {
"arxiv_id": "2206.14467",
"language": "zh",
"page_idx": 2
} | ||
2206.07277-zh-1 | 2206.07277-zh-1.jpg | # $\square\square$平滑的$\square$物：在$\square\square\square\square$下通$\square\square$助分$\square$器使用自适□□□平滑σ行□棒□□
Jongwoo Ko$^{1*}$, Bongsoo Yi$^2*$, Se-Young Yun$^1$
$^{1}$Kim Jaechul Graduate School of AI, KAIST, Seoul, Korea
$^{2}$Department of Statistics and Operations Research, University of North Carolina at Chapel ... | [
[
0.293212890625,
0.1256103515625,
0.70166015625,
0.12371826171875,
0.70166015625,
0.141357421875,
0.293212890625,
0.143310546875
],
[
0.39208984375,
0.14453125,
0.61279296875,
0.14453125,
0.61279296875,
0.1622314453125,
0.39208984375,
0.1622314453... | [
"平滑的物：在下通助分器使用",
"自适平滑行棒",
"Jongwoo Kol*, Bongsoo Yi2*, Se-Young Yunl",
"1 Kim Jaechul Graduate School of AI, KAIST, Seoul, Korea",
"2 Department of Statistics and Operations Research, University of North Carolina at Chapel Hill",
"(jongwoo.ko, yunseyoung @kaist.ac.kr, bongsoo@unc.edu",
"Abstract",
"... | {
"arxiv_id": "2206.07277",
"language": "zh",
"page_idx": 1
} | ||
2104.05458-zh-4 | 2104.05458-zh-4.jpg | <FigureHere>
Figure 3: The pipeline of PGNet: 1) Extract feature from an input image, and learn TCL, TBO, TDO, TCC maps as a multi-task problem; 2) The detection and recognition of each text instance can be achieved in a single shot by polygon restoration and PG-CTC decoding mechanism with the center point sequence of... | [
[
0.10455322265625,
0.07891845703125,
0.26708984375,
0.07891845703125,
0.26708984375,
0.0921630859375,
0.10455322265625,
0.0921630859375
],
[
0.39453125,
0.07574462890625,
0.44677734375,
0.0694580078125,
0.449462890625,
0.07891845703125,
0.39794921875,... | [
"Input Image:hxwx3",
"爱",
"Detection Branch",
"AND",
"1",
"TBO",
"hw",
"AND",
"4××4",
"um",
"InstanceInfo",
"RANDYS",
"S",
"TCL",
"CC",
"DONUTS",
"DONUTS",
"hw",
"DONUTS",
"**x1",
"/Center Point Seq",
"TDO",
"xx",
"Fsemantic",
"皖AM3C",
"RR-A-N-DDY-SS",
"Fvisual",
... | {
"arxiv_id": "2104.05458",
"language": "zh",
"page_idx": 4
} | ||
2006.09503-zh-0 | 2006.09503-zh-0.jpg | Memory-Efficient Pipeline-Parallel DNN Training
Deepak Narayanan$^1^*$ Amar Phanishayee$^2\text{ Kaiyu Shi}^3$ Xie Chen$^3\text{ Matei Zaharia}^1$
???。3in
Abstract
资源利用率不足E(Narayanan et al.,2019)或仅在多GPU 服务器E(Shoeybi et al.,2019) 内具有良好可扩展性的高通信开销，从而增加了训练时间和美元成本最近的工作提出了流水线模型并行性，以加速模型并行训练。例如，GPipe (Huang et al.,2019)和P... | [
[
0.23779296875,
0.11236572265625,
0.736328125,
0.11297607421875,
0.736328125,
0.1307373046875,
0.23779296875,
0.1300048828125
],
[
0.1732177734375,
0.1787109375,
0.79736328125,
0.1787109375,
0.79736328125,
0.1932373046875,
0.1732177734375,
0.19323... | [
"Memory-Efficient Pipeline-Parallel DNN Training",
"Deepak Narayanan 1 *Amar Phanishayee2Kaiyu Shi 3Xie Chen 3 Matei Zaharia 1",
"???。3in",
"Abstract",
"资源利用率不足E(Narayananetal.，2019）或仅在多",
"Many state-of-the-art ML results have been",
"GPU服务器E(Shoeybietal.，2019）内具有良好可扩",
"obtained by scaling up the nu... | {
"arxiv_id": "2006.09503",
"language": "zh",
"page_idx": 0
} | ||
2006.09503-zh-1 | 2006.09503-zh-1.jpg | Title Suppressed Due to Excessive Size
权重版本(称为阴影版本)用于已经在运行的输入。每个工作线程的权重的双重缓冲产生了一个流水线方案，其吞吐量高于 GPipe (没有管道刷新),并且内存效率高于 PipeDream (2 个权重版本，而深度为$d$ 的PipeDream 中的$d$的最坏情况)。2BW 在更新权重$(W^{(t+1)}=W^{(t)}-\nu\cdot\nabla f(W^{(t-1)})$的权重更新方程)时引入了一个恒定的权重延迟项 1, 在各个阶段之间保持一致，我们证明了该模型具有与香草权重更新 (§5.1 ) 经验上相似的模型收敛性。我们还提出了 2BW 的一个变体 (... | [
[
0.358642578125,
0.058074951171875,
0.61669921875,
0.058074951171875,
0.61669921875,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.09234619140625,
0.08709716796875,
0.4794921875,
0.08709716796875,
0.4794921875,
0.10101318359375,
0.09234619... | [
"Title Suppressed Due to Excessive Size",
"而是将新的权重用于新进人管道的输人，而将先前的",
"任何时间点最多只有一个工作线程处于活动状态。张",
"权重版本（称为阴影版本）用于已经在运行的输人。",
"量（层内）模型并行性（Shoeybietal，2019）导致关",
"每个工作线程的权重的双重缓冲产生了一个流水线方",
"键路径中昂贵的all-to-all通信，将模型分区的数量限",
"案，其吞吐量高于GPipe（没有管道刷新），并且内存",
"制为单个服务器中的GPU数量。FlexFlow（Jiaetal.",
"效率高于PipeDream... | {
"arxiv_id": "2006.09503",
"language": "zh",
"page_idx": 1
} | ||
2006.09503-zh-2 | 2006.09503-zh-2.jpg | Title Suppressed Due to Excessive Size
<FigureHere>
( b) PipeDream.
Figure 1. Timelines of different pipeline-parallel executions. Without loss of generality, forward and backward passes are assumed to take twice as long as forward passes; forward passes are shown in blue and backward passes are shown in green. Num... | [
[
0.35693359375,
0.05682373046875,
0.61767578125,
0.05682373046875,
0.61767578125,
0.0706787109375,
0.35693359375,
0.0706787109375
],
[
0.71337890625,
0.085205078125,
0.86767578125,
0.083984375,
0.86767578125,
0.0985107421875,
0.71337890625,
0.0997... | [
"Title Suppressed Due to Excessive Size",
"Before: w(1),w(2) w(3), w(4)",
"Pipeline",
"After: w(2), w(3), w(4), w(5)",
"flush",
"Operationsuseweightversion from lastflush",
"Worker1",
"123",
"26",
"537485",
"5",
"1",
"Worker1",
"123",
"3",
":",
"Worker2",
"2",
"1",
"2",
"3",
... | {
"arxiv_id": "2006.09503",
"language": "zh",
"page_idx": 2
} | ||
2006.09503-zh-3 | 2006.09503-zh-3.jpg | Title Suppressed Due to Excessive Size
$m=d$ (图 2 中的$d=4$ )。不能立即使用新的重量版本。特别是，飞行中的输入不能使用最新的重量版本进行向后传递(例如，在$t=21$的工作人员 3 上输人 7), 因为这些输入的前向传递已经在不同阶段使用较旧的重量版本启动。因此，新生成的权重版本需要缓冲以备将来使用。但是，需要维护的权重版本总数最多是 2,因为用于生成新权重版本的权重版本可以立即被丢弃(通过该阶段的未来输入不再使用旧权重版本)。例如，在图 2 中，每个工人可以丢弃$W_i^{(0)}$ 一旦他们完成处理输入 8 的向后传递，因为所有后续输入都使用较晚的权重版本进行向前和向后传... | [
[
0.358642578125,
0.058074951171875,
0.615234375,
0.058074951171875,
0.615234375,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.09234619140625,
0.08709716796875,
0.4755859375,
0.08709716796875,
0.4755859375,
0.10101318359375,
0.092346191406... | [
"Title Suppressed Due to Excessive Size",
"m=d（图2中的d=4）。不能立即使用新的重量",
"直观上，模型的参数在单次迭代中不会发生显著变化",
"版本。特别是，飞行中的输人不能使用最新的重量版",
"因此W(t)～W(t-1)。复制因子大于1的语义是相似",
"本进行向后传递（例如，在t=21的工作人员3上输",
"的，批处理大小乘以副本数量（与常规数据并行性一",
"人7），因为这些输入的前向传递已经在不同阶段使用",
"样）。可以类似地分析其他基于动量的优化器，例如",
"较旧的重量版本启动。因此，新生成的权重版本需要",
"Adam... | {
"arxiv_id": "2006.09503",
"language": "zh",
"page_idx": 3
} | ||
2006.09503-zh-4 | 2006.09503-zh-4.jpg | Title Suppressed Due to Excessive Size
4. Planner
道配置的缩减搜索空间上进行穷举搜索，确定如何在可用的计算设备上拆分模型。规划器还确定是否应部署节省内存的优化，以及每个 GPU 的微批处理大小和梯度度 accumulation , 给定最大安全全局批处理大小，验证不影响模型收敛(例如。g。,由过去的超参数扫描确定，无需流水线)
PipeDream-2BW 的计划程序使用成本模型来计算模型中各个块的计算时间和内存占用。时间和内存成本函数允许 PipeDream-2BW 推理管道宽度/深度和内存节省优化(例如激活重新计算)对吞吐量和内存占用的影响。例如，更深的配置具有额外的内存容量，... | [
[
0.358642578125,
0.058074951171875,
0.61669921875,
0.058074951171875,
0.61669921875,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.08905029296875,
0.08709716796875,
0.181396484375,
0.08709716796875,
0.181396484375,
0.0985107421875,
0.08905... | [
"Title Suppressed Due to Excessive Size",
"4.Planner",
"4.2. Partitioning Algorithm",
"PipeDream-2BW的规划器通过在所有可能的并行管",
"将它们放在一起，给定总内存容量M，PipeDream-",
"道配置的缩减搜索空间上进行穷举搜索，确定如何在",
"2BW的规划器首先确定适合给定工作程序的最大每",
"可用的计算设备上拆分模型。规划器还确定是否应部",
"GPU微批处理大小（以及相应的吞吐量），无论是否使",
"署节省内存的优化，以及每个GPU的微批处理大小",
"用内存成本函数部署... | {
"arxiv_id": "2006.09503",
"language": "zh",
"page_idx": 4
} | ||
2108.04927-zh-1 | 2108.04927-zh-1.jpg | $\textbf{Embodied BERT: 用 于 具 体 的 , 语 言 指 导 的 视 觉 任 务 完 成 的 转 换 器 模 型 }$
Alessandro Suglia$^1$
$Q\text{iaozi Gao}^{2}$
Jesse Thomason$^{2,3}$
Govind Thattai$^2$
Gaurav S. Sukhatme$^{2,3}$
$^{1}$Heriot-Watt University; $^2$Amazon Alexa AI; $^3$University of Southern California
$\mathbf{Abstract}$
Language-guided ro... | [
[
0.1234130859375,
0.0955810546875,
0.875,
0.0955810546875,
0.875,
0.10986328125,
0.1234130859375,
0.10986328125
],
[
0.16796875,
0.127685546875,
0.345947265625,
0.1270751953125,
0.345947265625,
0.144287109375,
0.16796875,
0.1455078125
],
[
... | [
"EmbodiedBERT：用于具体的，语言指导的视觉任务完成的转换器模型",
"Alessandro Suglial *",
"Qiaozi Gao?",
"Jesse Thomason?.3",
"Govind Thattai?",
"Gaurav S. Sukhatme2,3",
"1Heriot-Watt University:2Amazon Alexa AI; 3University of Southern California",
"Abstract",
"点是识别物体地标E(Blukisetal.2018）和目",
"的地E(Batraetal.，2020b）。但是，对于代理... | {
"arxiv_id": "2108.04927",
"language": "zh",
"page_idx": 1
} | ||
2108.04927-zh-2 | 2108.04927-zh-2.jpg | <FigureHere>
Figure 1: Embodied BERT. EmBERT attends to object detections in a panoramic view around an agent, then predicts an action and both a target object and target object parent for both navigation and manipulation actions. For example, at timesteps $t=0,1$ above, the model must predict the sink object target a... | [
[
0.480224609375,
0.08074951171875,
0.5224609375,
0.0831298828125,
0.5205078125,
0.09320068359375,
0.478515625,
0.0914306640625
],
[
0.5634765625,
0.08197021484375,
0.61279296875,
0.08197021484375,
0.61279296875,
0.0926513671875,
0.5634765625,
0.09... | [
"Step Inst",
"EmBERT",
"Back",
"Left",
"Right",
"Action",
"Goal Inst",
"Front",
"Obje",
"tBBoxPare",
"",
"E",
"Head to th",
"",
"sink in the",
"the sink",
"corner",
"Tert",
"Pouced",
"食",
"",
"",
"sinkinthe",
"Head to the",
".0",
"corner.",
"orwaro",
"Puta",
"... | {
"arxiv_id": "2108.04927",
"language": "zh",
"page_idx": 2
} | ||
2108.04927-zh-3 | 2108.04927-zh-3.jpg | <table>
<tbody>
<tr>
<th> </th>
<th colspan="2">$\mathbf{Language}$ $0bs.$</th>
<th colspan="2">$\mathbf{Visual}0\mathbf{bs}.$</th>
<th colspan="2">Historical $0bs.$ T</th>
<th colspan="2">Inference</th>
</tr>
<tr>
<th> </th>
<th>Goal Inst. Structure</th>
<th>Inst. Split</th>
<th>Views... | [
[
0.349365234375,
0.1561279296875,
0.44921875,
0.1561279296875,
0.44921875,
0.16748046875,
0.349365234375,
0.16748046875
],
[
0.491943359375,
0.1561279296875,
0.56689453125,
0.1561279296875,
0.56689453125,
0.16748046875,
0.491943359375,
0.167480468... | [
"Language Obs.",
"Visual Obs.",
"Historical Obs.",
"Inference",
"Goal Inst.",
"Inst.",
"As",
"Hidden",
"Mask",
"Nav Obj.",
"Views",
"Features",
"Structure",
"Split",
"Inputs",
"States",
"Pred.",
"Pred.",
"+",
"LSTM",
"Direct",
"x",
"SEQ2SEQ(Shridharetal.,2020)",
"x",
... | {
"arxiv_id": "2108.04927",
"language": "zh",
"page_idx": 3
} | ||
2108.04927-zh-4 | 2108.04927-zh-4.jpg | 。每个任务都涉及一个或多个需要操作的对象， 例如 apple ,以及它们应放在其上的最终容器，例如 plate。许多任务涉及中间状态更改，例如 HEAT & PLACE 需要在 microwave 中烹饪目标对象
监督数据。每个 ALFRED 情节包括模拟房间的初始状态，语言说明，计划目标和专家演示轨迹。语言指令作为高级目标指令$\vec{I}$上的 X$\vec{I}$, 例如将煮熟的鸡蛋放在水槽中，再加上一系列逐步的指令$\vec{I}$,例如右转并转到水槽右侧的柜台上的鸡蛋，...计划目标$\mathcal{P}$ (或子目标)是目标和参数的元组$\vec{\mathcal{I}}$,例如。专家演示轨迹$\mathcal{T}... | [
[
0.124267578125,
0.0885009765625,
0.5087890625,
0.0885009765625,
0.5087890625,
0.09857177734375,
0.124267578125,
0.09857177734375
],
[
0.51123046875,
0.0872802734375,
0.6826171875,
0.0872802734375,
0.6826171875,
0.1009521484375,
0.51123046875,
0.1... | [
"每个任务都涉及一个或多个需要操作的对象，",
"进行了比较（图E2）",
"例如apple，以及它们应放在其上的最终容",
"4EmbodiedBERT",
"器，例如plate。许多任务涉及中间状态更",
"改，例如HEAT&PLACE需要在microwave",
"EmBERT使用转换器编码器来共同嵌人语言",
"中烹饪目标对象",
"和视觉标记，并使用转换器解码器来进行长期",
"监督数据。每个ALFRED情节包括模拟房间",
"规划和以对象为中心的导航预测（图E3）",
"的初始状态，语言说明，计划目标和专家演示",
"轨迹。语言指令作为高级目标指令工上的X工",
... | {
"arxiv_id": "2108.04927",
"language": "zh",
"page_idx": 4
} | ||
1506.06863-zh-0 | 1506.06863-zh-0.jpg | # $\Delta$Bleu:生成任务的判别度量，具有本质上不同的目标$^*$
Michel Galley$^1\dagger$ Chris Brockett$^1$
Alessandro Sordoni$^2\sharp$ Yangfeng Ji$^3\sharp$
$\mathrm{Bill~Dolan}^{1}$
Michael Auli$^{4\ddagger\:\text{Chris Quirk}^1\:\text{Margaret Mitchell}^1\:\text{Jianfeng Gao}^1}$
$^{1}$Microsoft Research, Redmond, WA, USA
$^{2}$DIRO, Un... | [
[
0.403076171875,
0.0831298828125,
0.59423828125,
0.0843505859375,
0.59423828125,
0.1015625,
0.403076171875,
0.1009521484375
],
[
0.306396484375,
0.1051025390625,
0.7060546875,
0.1051025390625,
0.7060546875,
0.119384765625,
0.306396484375,
0.119384... | [
"△Bleu：生成任务",
"的判别度量，具有本质上不同的目标*",
"Michel Galley1t",
"ChrisBrockett1Alessandro Sordoni2",
"Yangfeng Ji3t",
"Michael Auli4t",
"",
"Chris QuirklMargaret MitchelllJianfeng Gaol",
"Bill Dolan",
"",
"",
"1Microsoft Research, Redmond, WA, USA",
"2DIRO, Universite de Montréal, Montreal, QC, Canada... | {
"arxiv_id": "1506.06863",
"language": "zh",
"page_idx": 0
} | ||
1506.06863-zh-1 | 1506.06863-zh-1.jpg | <FigureHere>
Figure 1: Example of consecutive utterances of a dia-
$\log.$
测试集上的索引。在 Bleu,$^{1}$的情况下，系统
输出$\underset{\mathrm{2cm}}{\operatorname*{h_{1}\ldots}}h_{I}$的自动得分定义为：
$$\mathrm{Bleu}=\mathrm{BP}\cdot\exp\left(\sum_n\log p_n\right)$$
(1)
。5cm ,带： ${}^{\text{“”}}5$cm
(2)
$$\mathrm{BP}=\begin{cases}\:1&\mathrm... | [
[
0.225830078125,
0.0748291015625,
0.3359375,
0.07659912109375,
0.3359375,
0.0914306640625,
0.225830078125,
0.08905029296875
],
[
0.54833984375,
0.07598876953125,
0.72314453125,
0.07598876953125,
0.72314453125,
0.09027099609375,
0.54833984375,
0.09... | [
"I'm on my way now.",
"Discriminative Bleu",
"3",
"Context",
"Message",
"ll be downstairs waiting",
"判别式Bleu或△Bleu通过利用与引用ri,",
"Response",
"I'll keep an eye out for you.",
"相关的人类定性判断wiiE【-1,+1]来扩展",
"Bleu。它是有区别的，因为它既奖励具有“良",
"Figure l: Example of consecutive utterances of a dia-",
"好”参考响应的匹配... | {
"arxiv_id": "1506.06863",
"language": "zh",
"page_idx": 1
} | ||
1506.06863-zh-3 | 1506.06863-zh-3.jpg | 制回来一样)。然后，由 5 名众包评分者根据上下文和信息对新的三元组的响应质量进行评分，每个评分者采用 5 分的 Likert 型量表。$^{3}$至关重要，再次与 Sordoni 等人形成对比。(2015) , 我们没有对这些合成多参考集施加分数截止值。相反，我们保留了所有候选答案， 并将其分数缩放为[-1,+1]
Table 1 presents representative multireference examples (from the dev set) together with their converted scores. The context and messages associated with the s... | [
[
0.12176513671875,
0.07659912109375,
0.48779296875,
0.07659912109375,
0.48779296875,
0.09027099609375,
0.12176513671875,
0.09027099609375
],
[
0.515625,
0.07598876953125,
0.87744140625,
0.07598876953125,
0.87744140625,
0.09027099609375,
0.515625,
... | [
"制回来一样）。然后，由5名众包评分者根据",
"长度比以及DSSM得分（Huangetal，2013)",
"上下文和信息对新的三元组的响应质量进行",
"RNN-based MT：根据(Sordoniet al.,2015)",
"评分，每个评分者采用5分的Likert型量表。",
"的RNN模型的对数概率。",
"3至关重要，再次与Sordoni等人形成对比。",
"Baseline：随机基线",
"(2015），我们没有对这些合成多参考集施加分",
"虽然△Bleu依赖于人类的定性判断，但重",
"数截止值。相反，我们保留了所有候选答案，",
"要的是要注意，人类对多参考（... | {
"arxiv_id": "1506.06863",
"language": "zh",
"page_idx": 3
} | ||
1506.06863-zh-4 | 1506.06863-zh-4.jpg | <table>
<tbody>
<tr>
<th>Metric</th>
<th>refs. </th>
<th>Spearman</th>
<th>Kendall</th>
<th>S $T$ </th>
</tr>
<tr>
<td>Bleu</td>
<td>$\operatorname{single}$</td>
<td>(.178, .260</td>
<td>.171 (.087.</td>
<td>.252</td>
</tr>
<tr>
<td>Bleu</td>
<td>0.6 $w$</td>
<td>.343 (... | [
[
0.26025390625,
0.07598876953125,
0.35693359375,
0.07781982421875,
0.35693359375,
0.0926513671875,
0.26025390625,
0.09027099609375
],
[
0.5458984375,
0.0706787109375,
0.5693359375,
0.0748291015625,
0.56494140625,
0.0872802734375,
0.54150390625,
0.... | [
"Spearman's p",
"0.6",
"Metric",
"refs.",
"Kendall's T",
"Bleu",
"0.5",
"single",
"260(.178,.337)",
".171 (.087,.252)",
"Bleu",
"w ≥ 0.6",
".343 (.265,.416)",
".232 (.150, .312)",
"nC",
"0.4",
".318 (.239, .392)",
"S",
"all",
"Bleu",
"212 (.129,",
"，,.292)",
"",
"0.3",
... | {
"arxiv_id": "1506.06863",
"language": "zh",
"page_idx": 4
} | ||
2006.11077-zh-0 | 2006.11077-zh-0.jpg | Published as a conference paper at ICLR 2021
一个更好的替代错误反馈的通信高效分布式学习
Samuel Horváth and Peter Richtárik
## KAUST
Thuwal, Saudi Arabia
$\{ samuel.horvath, peter.richtarik \} @kaust.edu.sa$
# ABSTRACT
Modern large-scale machine learning applications require stochastic optimization algorithms to be implemented on distr... | [
[
0.1748046875,
0.034088134765625,
0.47802734375,
0.034088134765625,
0.47802734375,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.1968994140625,
0.10418701171875,
0.76123046875,
0.10418701171875,
0.76123046875,
0.125,
0.1968994140625,
0... | [
"Published as a conference paper at ICLR 2021",
"个更好的替代错误反馈的通信高效分布式学习",
"Samuel Horvath and PeterRichtarik",
"KAUST",
"Thuwal, SaudiArabia",
"f samuel.horvath, peter.richtarik l @kaust.edu.sa",
"ABSTRACT",
"Modern large-scale machine learning applications require stochastic optimization",
"algorithm... | {
"arxiv_id": "2006.11077",
"language": "zh",
"page_idx": 0
} | ||
2006.11077-zh-1 | 2006.11077-zh-1.jpg | $\underline{\text{Published as a conference paper at ICLR 2021}}$
组合，以获得更积极的效果。我们将不再进一步区分稀疏化和量化方法，并将它们全部称为
压缩运算符
$\bullet$ Induced Compressor. 在稳定 EF 框架内使用时，偏置压缩机 (e。g。, Top- $K$ )与无偏见的
。通过随机稀疏化技术可以获得许多稀疏向量，这些稀疏化技术可以随机掩盖输人向量，并且仅保留恒定数量的坐标$\mathbb{E}(????)$。还有一个工作线$\mathbb{E}(??)$, 其中提出了稀疏化和量化的考虑到实践和理论，压缩运算符可以分为两组：有偏和... | [
[
0.1748046875,
0.034088134765625,
0.476318359375,
0.034088134765625,
0.476318359375,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.1895751953125,
0.106689453125,
0.82177734375,
0.106689453125,
0.82177734375,
0.117431640625,
0.1895751953125... | [
"Published as a conference paper at ICLR 2021",
"通过随机稀疏化技术可以获得许多稀疏向量，这些稀疏化技术可以随机掩盖输人向量，并",
"且仅保留恒定数量的坐标E（?????）。还有一个工作线E（?？），其中提出了稀疏化和量化的",
"组合，以获得更积极的效果。我们将不再进一步区分稀疏化和量化方法，并将它们全部称为",
"压缩运算符",
"考虑到实践和理论，压缩运算符可以分为两组：有偏和无偏。对于无偏压缩器，要求C（9）是",
"更新9的无偏估计器。一旦这一要求被解除，额外的技巧是必要的分布式压缩随机梯度下降",
"(DCSGD）（???？）采用这样的... | {
"arxiv_id": "2006.11077",
"language": "zh",
"page_idx": 1
} | ||
2006.11077-zh-2 | 2006.11077-zh-2.jpg | Published as a conference paper at ICLR 2021
<table>
<tbody>
<tr>
<th>$\overline{\mathbf{Algo}}$</th>
<th>rithm1DCSGD</th>
<th>gorithm 2 DCSGD with Error Feedback</th>
</tr>
<tr>
<td>1:</td>
<td>$\overline{[\mathbf{nput:}\{\eta^{k}\}_{k=0}^{T}>0\:,x_{0}]}$</td>
<td>=0 $\forall i\in [ n]$ $\mat... | [
[
0.1732177734375,
0.0328369140625,
0.47802734375,
0.032196044921875,
0.47802734375,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.176513671875,
0.11932373046875,
0.32177734375,
0.11932373046875,
0.32177734375,
0.1331787109375,
0.17651367187... | [
"Published as a conference paper at ICLR 2021",
"Algorithm 1DCSGD",
"Algorithm2DCSGDwithErrorFeedback",
"1: Input: (n=0 ≥ 0, o",
"1: Input: (n')T=o > 0, ao,e = 0 Vi e [n]",
"2: for k =0, 1,...T do",
"2: for k=0.1.",
"...T do",
"Parallel: Worker side",
"Parallel: Worker side",
"3:",
"3:",
"4:... | {
"arxiv_id": "2006.11077",
"language": "zh",
"page_idx": 2
} | ||
2006.11077-zh-3 | 2006.11077-zh-3.jpg | $\underline{\text{Published as a conference paper at ICLR 2021}}$
如果$\delta=1$ (无压缩),定理$\overline{\mathrm{E}}$2 恢复分布式 SGD$\overline{\mathrm{E}}(?)$的最佳速率。如果$\delta>1$,则收敛速度中有一个额外的项$(\delta_n-1)D$ ,这是由于数据的异质性而出现的$(\sum_i=1^n\nabla f_i(x^\star)=0$,但通常是$\sum_i=1^n\mathcal{C}(\nabla f_i(x^\star))\neq0$ )。此外，由于存在导致$L\to \de... | [
[
0.1748046875,
0.034088134765625,
0.476318359375,
0.034088134765625,
0.476318359375,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.17724609375,
0.10479736328125,
0.82177734375,
0.10479736328125,
0.82177734375,
0.11871337890625,
0.177246093... | [
"Published as a conference paper at ICLR 2021",
"如果=1（无压缩），定理E2恢复分布式SGDE(?）的最佳速率。如果>1，则收敛速",
"度中有一个额外的项（8n-1)D，这是由于数据的异质性而出现的（Z=1Vf:(c*)=0，但",
"通常是-,C(Vf(α*))0）。此外，由于存在导致L→onL和/n→/n的压缩，速",
"率受到额外方差的负面影响",
"接下来，我们将我们的速率与与CEU()CC()",
"[(T)-] =(Lrexp[-] +)",
"LT",
"",
"一起使用的错误反馈E(?）E（n=1），（?）E（n≥1）的最... | {
"arxiv_id": "2006.11077",
"language": "zh",
"page_idx": 3
} | ||
2006.11077-zh-4 | 2006.11077-zh-4.jpg | $\underline{\text{Published as a conference paper at ICLR 2021}}$
<FigureHere>
Figure 1: Comparison of Top- 1 (+ EF) and NU Rand- 1 on Example 1 from ?
$\delta_3=\delta-\left(1-\frac1\delta\right)<\delta.$ 当传输消息的大小加倍时，可以使用算法$\mathbb{E}$1,因为$\mathcal{C}$ 是无偏的，它提供了比算法更好的收敛保证$\overline{\mathrm{E}}$2。基于诱导压缩器的构造，人们可能会期望我... | [
[
0.176513671875,
0.034088134765625,
0.476318359375,
0.034088134765625,
0.476318359375,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.57421875,
0.111083984375,
0.6005859375,
0.111083984375,
0.6005859375,
0.11871337890625,
0.57421875,
0.118713... | [
"Published as a conference paper at ICLR 2021",
"NURar",
"Top-1 + EF",
"Fop-1+E",
"Top-",
"101",
"on",
"AA",
"10",
"10(",
"160",
"10",
"20",
"60",
"20",
"80",
"",
"10",
"Figure 1: Comparison of Top- 1 (+ EF) and NU Rand- 1 on Example 1 from ?",
"83=8-（1-）<8.当传输消息的大小加倍时，可以使用算法E1... | {
"arxiv_id": "2006.11077",
"language": "zh",
"page_idx": 4
} | ||
2003.11523-zh-0 | 2003.11523-zh-0.jpg | $\underline{\text{Published as a conference paper at ICLR 2020}}$
具有迁移学习的 Tigrinya 神经机器翻译用于人道主义
# 响应
$\operatorname*{Alp}_{\mathrm{T}}$Öktem, Mirko Plitt $\&$ Grace Tang
Translators without Borders
30 Main Street STE 500 $\mathop{\mathrm{Danbury,~CT~06810,~USA}}$ $\{$ alp,mirko,grace $\}$ @translatorswithoutborders.o... | [
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.1805419921875,
0.10479736328125,
0.81884765625,
0.10479736328125,
0.81884765625,
0.12249755859375,
0.18054... | [
"Published as a conference paper at ICLR 2020",
"具有迁移学习的Tigrinya神经机器翻译用于人道主义",
"响应",
"AlpOktem,MirkoPlitt&GraceTang",
"TranslatorswithoutBorders",
"30 Main Street STE 500",
"Danbury, CT 06810, USA",
"[alp,mirko,grace@translatorswithoutborders.org",
"Abstract",
"We report our experiments in buildin... | {
"arxiv_id": "2003.11523",
"language": "zh",
"page_idx": 0
} | ||
2003.11523-zh-1 | 2003.11523-zh-1.jpg | $\underline{\text{Published as a conference paper at ICLR 2020}}$
Table 1: Parallel corpora used in this work. Dataset names marked with an asterisk are available through OPUS repository (Tiedemann, 2012; Christodouloupoulos & Steedman, 2015) . Ethiopian languages corpus (Teferra Abate et al., 2018) is also openly ava... | [
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.1748046875,
0.11297607421875,
0.82421875,
0.11297607421875,
0.82421875,
0.126953125,
0.1748046875,
0.126953125... | [
"Published as a conference paper at ICLR 2020",
"Table 1:Parallel corpora used in this work. Dataset names marked with an asterisk are",
"available through OPUS repository (Tiedemann, 2012; Christodouloupoulos & Steedman",
"2015) : Ethiopian languages corpus (Teferra Abate et al., 2018) is also openly availab... | {
"arxiv_id": "2003.11523",
"language": "zh",
"page_idx": 1
} | ||
2003.11523-zh-2 | 2003.11523-zh-2.jpg | $\underline{\text{Published as a conference paper at ICLR 2020}}$
References
$\text{\v {Z}eljko Agi\' {c}andIvanVuli\' {c}. JW300: Awide- coverage parallel corpus for low- resource lan- }$
guages. In Proceedings of the 57th Annual Meeting of the Association for Computational
Linguistics, pp. 3204-3210, Florence, Ital... | [
[
0.1748046875,
0.0328369140625,
0.5087890625,
0.0328369140625,
0.5087890625,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.1732177734375,
0.10418701171875,
0.2646484375,
0.10418701171875,
0.2646484375,
0.11871337890625,
0.1732177734375,
0.1187... | [
"Published as a conference paper at ICLR 2020",
"References",
"Zeljko Agic and Ivan Vulic. JW3oo: A wide-coverage parallel corpus for low-resource lan",
"guages. In Proceedings of the 57th Annual Meeting of the Association for Computational",
"Linguistics, pp. 3204-3210, Florence, Italy, July 2019.",
"Dzm... | {
"arxiv_id": "2003.11523",
"language": "zh",
"page_idx": 2
} | ||
2003.11523-zh-3 | 2003.11523-zh-3.jpg | $\underline{\text{Published as a conference paper at ICLR 2020}}$
Yemane Tedla and Kazuhide Yamamoto. The effect of shallow segmentation on englishtigrinya statistical machine translation. 2016 International Conference on Asian Language ${\mathrm{Processing~(IALP),~pp.~79-82,~2016.}}$
Solomon Teferra Abate, Michael Me... | [
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.176513671875,
0.10479736328125,
0.8203125,
0.10479736328125,
0.8203125,
0.11932373046875,
0.176513671875,
0.11... | [
"Published as a conference paper at ICLR 2020",
"Yemane Tedla and Kazuhide Yamamoto. The effect of shallow segmentation on english.",
"tigrinya statistical machine translation. 2016 International Conference on Asian Language",
"Processing (IALP), Pp.79-82, 2016.",
"Solomon Teferra Abate, Michael Melese, Mar... | {
"arxiv_id": "2003.11523",
"language": "zh",
"page_idx": 3
} | ||
2310.11163-zh-1 | 2310.11163-zh-1.jpg | $\mathbf{IMTL}_{\mathrm{AB}}:$用于构建、评估和诊断交互式机器翻译系统的开源平台
Xu Huang$^{1*}$ Zhirui Zhang$^2\textbf{ Ruize Gao}^3$ Yichao Du$^4$ Lemao Liu$^2$ $\bar{\textbf{Gouping Huang}^2}$ $\bar{\textbf{Shuming Shi}^2}$ Jiajun Chen$^1$ $\textbf{Shujian Huang}^{1\dagger }$
$^{1}$National Key Laboratory for Novel Software Technology, Nanj... | [
[
0.14697265625,
0.0872802734375,
0.8486328125,
0.0885009765625,
0.8486328125,
0.10272216796875,
0.14697265625,
0.1015625
],
[
0.19482421875,
0.11639404296875,
0.8134765625,
0.11639404296875,
0.8134765625,
0.1312255859375,
0.19482421875,
0.13122558... | [
"IMTLAB：用于构建、评估和诊断交互式机器翻译系统的开源平台",
"Xu Huangl*Zhirui Zhang?Ruize Gao3Yichao Du4 Lemao Liu?",
"Gouping Huang?Shuming Shi?Jiajun Chen!Shujian Huangl",
"1National Key Laboratory for Novel Software Technology, Nanjing University",
"2Tencent AI Lab3Shanghai Jiao Tong University",
"4University of Science and Te... | {
"arxiv_id": "2310.11163",
"language": "zh",
"page_idx": 1
} | ||
2310.11163-zh-3 | 2310.11163-zh-3.jpg | 系统的响应。此交互过程一直持续到满足用户 束
的目标或用户对系统感到沮丧为止。通过这种为干预来完善翻译，并使研究人员能够以端到端的方式评估不同的 IMT 系统
# 3.2 Communication Interface
为了支持各种 IMT 系统的灵活体系结构，我们设计了 IMT 系统与用户之间的通用通信接口。用户仅限于五种常见的编辑操作：(i)保持单词不变；(ii)在任何位置插入连续的单词跨度；(iii) 将连续的单词跨度替换为新的单词跨度；(iv)删除任何位置的连续单词跨度；(v) 将连续的单词跨度替换为特殊的占位符(或仅插人),以提示 IMT 系统填空。这五种常见类型可以组合在一起，以涵盖大多数 IMT 范式中的交互式策略... | [
[
0.12005615234375,
0.0872802734375,
0.484375,
0.0872802734375,
0.484375,
0.097412109375,
0.12005615234375,
0.097412109375
],
[
0.5146484375,
0.0885009765625,
0.87890625,
0.0885009765625,
0.87890625,
0.09857177734375,
0.5146484375,
0.09857177734375... | [
"根据用户的反馈生成翻译，用户提供编辑操作",
"作都可以转换为受词法约束的模板t，其中保",
"（例如保留，插人，替换，删除和填空）以请求",
"留，插入和替换的字符是IMT系统的词法约",
"系统的响应。此交互过程一直持续到满足用户",
"束",
"的目标或用户对系统感到沮丧为止。通过这种",
"请注意，我们当前的版本仅支持上述与键盘",
"方式，IMTLAB允许IMT系统利用明确的人",
"相关的操作，但它能够评估大多数常见的IMT",
"为于预来完善翻译，并使研究人员能够以端到",
"系统。我们将把适应留给其他特殊操作，如鼠",
"端的方式评估不同的IMT系统",
... | {
"arxiv_id": "2310.11163",
"language": "zh",
"page_idx": 3
} | ||
2310.11163-zh-4 | 2310.11163-zh-4.jpg | <FigureHere>
Figure 2: An example of the communication interface, where users perform editing operations (i.e., keep , insert , replace , delete and blank-filling ) on the output of the IMT system according to their own goals. We show two different editing processes from users: one of which contains a prefix constrain... | [
[
0.185546875,
0.0872802734375,
0.69287109375,
0.0872802734375,
0.69287109375,
0.1009521484375,
0.185546875,
0.1009521484375
],
[
0.160400390625,
0.1015625,
0.8310546875,
0.10394287109375,
0.8310546875,
0.1217041015625,
0.160400390625,
0.1193847656... | [
"Source：VR作品的最大优势是令观众具有独一无二的观影“沉浸\"感。",
"User Goal: The biggest advantage for VR works is to enable the audience to have a unique",
"movie watching \"immersion\".",
"IMT System: The greatest advantage of VR works is that the audience has a unique sense of",
"\"immersion\"in thefilm.",
".Human 1(prefix con... | {
"arxiv_id": "2310.11163",
"language": "zh",
"page_idx": 4
} | ||
2310.18830-zh-1 | 2310.18830-zh-1.jpg | ## 翻译掉没有并行数据的 Translitionese
Rricha Jalota$^{1,3,*},\textbf{Koel Dutta Chowdhury}^{1}$,
Cristina España-Bonet$^2$, and Josef van Genabith$^{1,2}$
$^{1}\bar{\text{Saarland University, Saarland Informatics Campus, Germany}}$ $^{2}$German Research Center for Artificial Intelligence (DFKI)
$^3$AppTek GmbH, Germany
rjalota... | [
[
0.306396484375,
0.08489990234375,
0.6943359375,
0.08489990234375,
0.6943359375,
0.09979248046875,
0.306396484375,
0.09979248046875
],
[
0.316650390625,
0.10809326171875,
0.68603515625,
0.10809326171875,
0.68603515625,
0.1258544921875,
0.316650390625,... | [
"翻译掉没有并行数据的Translitionese",
"Rricha Jalota',3, , Koel Dutta Chowdhury',",
"Cristina Espana-Bonet?, and Josef van Genabith12",
"I Saarland University, Saarland Informatics Campus, Germany",
"2German Research Center for Artificial Intelligence (DFK)",
"3AppTek GmbH, Germany",
"rjalota@apptek.com,rrja00001... | {
"arxiv_id": "2310.18830",
"language": "zh",
"page_idx": 1
} | ||
2310.18830-zh-2 | 2310.18830-zh-2.jpg | inal - translated dataset (in the same lan- 通过使用来自单独语言对$\overline{\mathrm{E}}(??)$ 的验证集或根
guage) during training and validation.
· We show that our method is able to reduce the accuracy of a translationse classifier to that of a random classifier, indicating that our approach is able to successfully eliminate translat... | [
[
0.15283203125,
0.08612060546875,
0.50537109375,
0.0872802734375,
0.50537109375,
0.1009521484375,
0.15283203125,
0.09979248046875
],
[
0.51318359375,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51318359375,
0.100... | [
"inal - translated dataset (in the same lan-",
"通过使用来自单独语言对E（??）的验证集或根",
"guage) during training and validation.",
"本不使用E(?）的验证方案，在NMT中强制执",
"行严格的无监督设置，从而冒着次优模型的风",
": We show that our method is able to reduce the",
"险。为了解决这个问题，？？？建议在单语数据",
"accuracy of a translationse classifier to that",
"上使用无监督验证... | {
"arxiv_id": "2310.18830",
"language": "zh",
"page_idx": 2
} | ||
2310.18830-zh-3 | 2310.18830-zh-3.jpg | <FigureHere>
Figure 1: Model architecture. Here $(tr,og)$ is a (Translated (source), Original (target)) style sentence pair, and $og$ -like is the translationese-mitigated output. The dashed arrows correspond to on-the-fly parallel pair extraction that facilitates Supervised Training, while the red arrows in bold repr... | [
[
0.3837890625,
0.10272216796875,
0.4013671875,
0.10272216796875,
0.4013671875,
0.11224365234375,
0.3837890625,
0.11224365234375
],
[
0.5439453125,
0.11102294921875,
0.60107421875,
0.11102294921875,
0.60107421875,
0.1181640625,
0.5439453125,
0.1181... | [
"",
"PretrainedLM",
"Cross Entropy",
"Loss",
"Decoder",
"(tr, og-like)",
"Semantic",
"Semantic Similarity",
"Similarity",
"Loss",
"Encod",
"AC",
"SPE",
"Figure 1: Model architecture. Here (tr , og ) is a (Translated (source), Original (target) style sentence pair, and og -like is the",
"... | {
"arxiv_id": "2310.18830",
"language": "zh",
"page_idx": 3
} | ||
2402.05515-zh-3 | 2402.05515-zh-3.jpg | <FigureHere>
Figure 3: The normalized label entropy $H_n^l$ on both models and 7 datasets with and without appropriatenoised NoısYICL. In most cases, the $H_n^l$ with NoIsYICL( w/ ) is greater than without NoısYICL( w/o ).
and Komatsuzaki, 2021) 。模型检查点从 hugging-
face $^2$加载
Hyperparameters. 我们将噪声分布的标准偏差$\sigma$固定为0.... | [
[
0.204833984375,
0.0831298828125,
0.2359619140625,
0.0831298828125,
0.2359619140625,
0.0914306640625,
0.204833984375,
0.0914306640625
],
[
0.397216796875,
0.0831298828125,
0.4248046875,
0.0831298828125,
0.4248046875,
0.0914306640625,
0.397216796875,
... | [
"GPT-2",
"GPT-",
": Lower prediction bias. When no valid query",
"中",
"Aan",
"60",
"is given, the predicted labels should have bal-",
"ancedfrequencies.However,predictionscon",
"ducted by under-calibrated LMs usually have",
"0",
"significant bias, which is harmful to ICL (Fei",
"旺",
"门",
"... | {
"arxiv_id": "2402.05515",
"language": "zh",
"page_idx": 3
} | ||
2402.05515-zh-4 | 2402.05515-zh-4.jpg | 这些带有“中性”标签的数据集(详见附录 A
)
结果如图所示。E3, 表明在大多数情况下， NoIsYICL 会校准整体偏差，尤其是在 GPT-2上
和
<FigureHere>
Figure 4: Left: Reliability diagrams (sparse bars) and global confidence distribution (dense bars) of GPT-2 on hate\_speech18 with( w/ $,ECE_{1}=9.01\%$ ) and without( w/o $,ECE_1=28.14\%)$ NoIsYICL. The predictions are divided... | [
[
0.118408203125,
0.08612060546875,
0.486083984375,
0.08612060546875,
0.486083984375,
0.09979248046875,
0.118408203125,
0.09979248046875
],
[
0.1368408203125,
0.119384765625,
0.48193359375,
0.119384765625,
0.48193359375,
0.133056640625,
0.1368408203125... | [
"这些带有“中性”标签的数据集（详见附录A",
"结果如图所示。E3，表明在大多数情况下",
"NOISYICL会校准整体偏差，尤其是在GPT-2",
"上",
"和",
"NoisYICL promotes faithful confidence",
"可以获得ExpectedCalibrationError（ECEp",
"(Naeinietal.，2015）是分类任务中置信度真实",
"00",
"Confidence of Predict",
"性的广泛使用的指标：",
"Figure 4: Left : Reliability diagrams (sparse bars)... | {
"arxiv_id": "2402.05515",
"language": "zh",
"page_idx": 4
} | ||
2210.07907-zh-2 | 2210.07907-zh-2.jpg | <FigureHere>
Figure 1: Illustration of using distance scores for poisoned sample detection. We attack a BERT (Devlin et al., 2019) model using the BadNet (Gu et al., 2017) method with a rare word trigger“mn”on the IMDB ( Maas et al., 2011) sentiment analysis task. Class 0 denotes negative and class l denotes positive.... | [
[
0.255126953125,
0.08612060546875,
0.420654296875,
0.08612060546875,
0.420654296875,
0.0968017578125,
0.255126953125,
0.0968017578125
],
[
0.587890625,
0.0843505859375,
0.8037109375,
0.08489990234375,
0.8037109375,
0.0968017578125,
0.587890625,
0.... | [
"UMAPProjectionofFeatures",
"Distribution of Mahalanobis Distances",
"Clean Class 0",
"0.0035",
"CleanSamples",
"Clean Class 1",
"12.5",
"",
"PoisonedSamples",
"Poisoned (0 →1)",
"0.003",
"",
"191",
"2.5",
"10",
"5.0",
"1000:",
"4000",
"5000",
"6000",
"UMAPDimI",
"Mahalanob... | {
"arxiv_id": "2210.07907",
"language": "zh",
"page_idx": 2
} | ||
2210.07907-zh-3 | 2210.07907-zh-3.jpg | 线后门防御方法来保护 NLP 模型。
# 3 Methodology
3.1 Preliminaries
Problem Setting 我们专注于用户缺乏从头
开始训练大型模型的能力并从不受信任的第三方获得预训练模型以用于进一步的个人目的的场景。用户可以直接部署受害者模型，也可以在部署之前在其小数据集上对其进行微调。但是，第三方可能是攻击者，并已将后门注入模型中。后门模型将在干净的数据上保持良好的性能，但是一旦在激活后门的输入中有触发器，它将始终预测目标标签。我们假设用户有一个重要的标签要保护 (例如。g。, 垃圾邮件分类系统中的非垃圾邮件类),这很可能与攻击者的目标标签相同$\overline{\mathrm{E}}($Y... | [
[
0.118408203125,
0.0872802734375,
0.387939453125,
0.0872802734375,
0.387939453125,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.51318359375,
0.08612060546875,
0.68994140625,
0.08489990234375,
0.68994140625,
0.09979248046875,
0.51318359375,
... | [
"线后门防御方法来保护NLP模型。",
"可以在干净验证集的1",
"3 Methodology",
"Q=Ng neDan",
"1f(a),",
"",
"3.1Preliminaries",
"2i=EsorePu",
" Z (f(r)-c)(f(n)-c)",
"1",
"ProblemSetting我们专注于用户缺乏从头",
"开始训练大型模型的能力并从不受信任的第",
"三方获得预训练模型以用于进一步的个人目的",
"上进行估计，其中XMATHX属于XMATHX",
"的场景。用户可以直接部署受害者模型，也可",
"类。我们使用任意Mahalanobi... | {
"arxiv_id": "2210.07907",
"language": "zh",
"page_idx": 3
} | ||
2212.06224-zh-2 | 2212.06224-zh-2.jpg | 一样有效。然而，宏观县域限制的中间策略——当县被智能分组时——可以平衡政策灵活性和有效性之间的权衡。我们证明，对于给定的空间粒度，找到最有效的县分区等效于解决由我们的溢出估计加权的无向县图上的最小$k$切割问题。使用这种方法，我们确定了宏观县的限制，这些限制恢复了全州范围内流动性减少的$90\%$,同时在县之间保持了很大的灵活性
总而言之，我们的贡献如下：
$\bullet$ Setting: we identify a novel setting for studying
spillovers where the same set of policies was applied
$\bar{\mathrm{with~the~sa... | [
[
0.1004638671875,
0.0706787109375,
0.398681640625,
0.0706787109375,
0.398681640625,
0.083984375,
0.1004638671875,
0.083984375
],
[
0.426513671875,
0.07135009765625,
0.476318359375,
0.07135009765625,
0.476318359375,
0.08209228515625,
0.426513671875,
... | [
"-样有效。然而，宏观县域限制的中间策略",
"当县被",
"(a) ",
"(b)",
"一可以平衡政策灵活性和有效性之间的",
"CO.",
"权衡。我们证明，对于给定的空间粒度，找到最有效的",
"智能分组时-",
"县分区等效于解决由我们的溢出估计加权的无向县图",
"上的最小k切割问题。使用这种方法，我们确定了宏",
"观县的限制，这些限制恢复了全州范围内流动性减少的",
"90%，同时在县之间保持了很大的灵活性",
"总而言之，我们的贡献如下：",
"(c)",
". Setting: we identify a novel setting for stud... | {
"arxiv_id": "2212.06224",
"language": "zh",
"page_idx": 2
} | ||
2212.06224-zh-3 | 2212.06224-zh-3.jpg | 日至 2021 年 1 月 25 日这段时间变灰，在此期间，该州的大部分地区都处于区域居家令 (CDPH 2020)下。我们可以看到，各县通常在相似的时间通过相似的层级， 这是可以预期的，因为 COVID-19 的严重程度在各县之间是相关的。然而，在许多星期里，我们也看到至少两个不同层次的实质性代表。例如，在 2021 年 3月 15日当周，紫层有 11 个县，红层有 42 个县，橙层有 4 个县，黄层有 1 个县 (图 1 b)。其中许多不同的层级出现在相邻的县之间，从而可以分析县边界之间的溢出效应Mobility network. 我们使用 SafeGraph 的数据，这是一家从移动应用程序匿名和聚合位置数据的公司。对于每个 P... | [
[
0.09234619140625,
0.07135009765625,
0.4755859375,
0.07135009765625,
0.4755859375,
0.08209228515625,
0.09234619140625,
0.08209228515625
],
[
0.52294921875,
0.07135009765625,
0.91015625,
0.07135009765625,
0.91015625,
0.08209228515625,
0.52294921875,
... | [
"日至2021年1月25日这段时间变灰，在此期间，该州",
"标准。对于一个大县（人口超过106.000），可以通过",
"的大部分地区都处于区域居家令（CDPH2020）下。我",
"两种方式满足从紫色到红色的标准：（1）通过满足所有",
"们可以看到，各县通常在相似的时间通过相似的层级",
"三个指标的红色层的阈值，（2）通过满足测试的阈值橙",
"这是可以预期的，因为COVID-19的严重程度在各县",
"色层的积极性和健康公平性，从而将调整后的病例率交",
"之间是相关的。然而在许多星期里，我们也看到至少",
"换为其他两个更严格的阈值。Small counties(popula-... | {
"arxiv_id": "2212.06224",
"language": "zh",
"page_idx": 3
} | ||
2311.05006-zh-1 | 2311.05006-zh-1.jpg | 1. False Familiarity vs. False Novelty: What type of
attack is more effective?
2. FGSM vs. iterative attacks: Is it worth exploring
more flexible iterative attacks to improve upon the
fast gradient sign method (FGSM)?
3. Uninformed vs. informed attacks: How can ad-
versarial attacks proft when the type of input is
gi... | [
[
0.0906982421875,
0.091552734375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.106689453125,
0.0906982421875,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.66650390625,
0.0921630859375,
0.66650390625,
0.1060791015625,
0.50244140625,
0.1... | [
"1. False Familiarity vs. False Novelty: What type of",
"2.3. Iterative attacks",
"attack is more effective?",
"送代方法可以产生更多不同的扰动，但计算成本更",
"2. FGSM vs. iterative attacks: Is it worth exploring",
"高。基本选代方法（BIM）【12]送代地应用FGSM更",
"more flexible iterative attacks to improve upon the",
"新，并通过以下方式进行描述：",
"fa... | {
"arxiv_id": "2311.05006",
"language": "zh",
"page_idx": 1
} | ||
2311.05006-zh-2 | 2311.05006-zh-2.jpg | <FigureHere>
Figure 2: Uninformed FGSM attacks. Fast gradient sign method (FGSM) attacks on TinyImageNet. Left: False
Familiarity (false negative, FN) attacks. Right: False Novelty (false positive, FP) attacks. (a,b) The OSR ranking measured by AUROC. ( c, d) Median Maximum Logit Score ( MLS) w. r. t. original scores... | [
[
0.101318359375,
0.10418701171875,
0.1258544921875,
0.10418701171875,
0.1258544921875,
0.11553955078125,
0.101318359375,
0.11553955078125
],
[
0.52685546875,
0.10479736328125,
0.54736328125,
0.10479736328125,
0.54736328125,
0.11553955078125,
0.5268554... | [
"0.8",
"0.8",
"2-Norm Loss",
"",
"Max Loss",
"Sum-ExpLoss",
"",
"Log-MSPLoss",
"Max Loss",
"0.5",
"2-Norm Loss",
"0.5",
"0.4",
"0.8",
"0.4",
"0.0",
"0.2",
"0.4",
"0.6",
"1.0",
"0.0",
"0.2",
"0.4",
"0.6",
"0.8",
"1.0",
"-Sizeof theAdvesarial Perturbation",
" - Si... | {
"arxiv_id": "2311.05006",
"language": "zh",
"page_idx": 2
} | ||
2311.05006-zh-3 | 2311.05006-zh-3.jpg | <FigureHere>
Figure 3: Informed attacks. False Positive (FP) attacks are performed on familiar samples (2-norm loss) and False Negative (FN) attacks on novel samples (max loss) . ( a) Fast gradient sign method ( FGSM) using $\varepsilon=0.07$ for FN and $\varepsilon=0.06$ for FP.(b) Our iterative method with $\varepsi... | [
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.40185546875,
0.10797119140625,
0.459228515625,
0.10797119140625,
0.459228515625,
0.11871337890625,
0.40185546875,
0.118713... | [
"terich&Guyer的[3预测，即可以通过对抗性扰动",
"AUROC:0.835",
"轻松提高logit得分。然而，这种能力只会导致在知情",
"AUR0C:0.835",
"No Attack",
"NoAttack",
"FP&FNAttack",
"AUR0C:0.807",
"",
"FP&FNAttackAUR0C:0.00",
"设置中有效的错误熟悉（FN）攻击。在不知情的情况",
"FPAttack",
"FPAttack",
"AUROC:8.340",
"AUROC;0.064",
"FNAttack",
"A0.8",
"AUR... | {
"arxiv_id": "2311.05006",
"language": "zh",
"page_idx": 3
} | ||
2310.00280-zh-0 | 2310.00280-zh-0.jpg | push0 g 0 G$\underline {\mathrm{Preprint}}$
$\therefore$pop
=1
MATHX\_THMAXXMAX\_2
$\_$THMAX\_3\_Corex\_Corex-Review$_Code$
$Corex:$通过多模型协作推动复杂推理的边界
$\underset{\sim}{\operatorname*{\operatorname*{\operatorname*{Qiushi~Sun}}}}^{\circ\circ\circ*},\overset{\circ\circ}{\operatorname*{Zhangyue~Yin}}^{\bullet},\overset{... | [
[
0.1732177734375,
0.037261962890625,
0.313720703125,
0.0309295654296875,
0.314453125,
0.048614501953125,
0.1748046875,
0.054931640625
],
[
0.90283203125,
0.039154052734375,
0.939453125,
0.039154052734375,
0.939453125,
0.054931640625,
0.90283203125,
... | [
"'pushO g O GPreprint",
"-pop",
"=1",
"MATHX_THMAXXMAX_2",
"THMAX_3_Corex-",
"Corex-ReviewCod",
"Corex：通过多模型协作推动复杂推理的边界",
"",
"Qiushi Sun *, Zhangyue Yint, Xiang Li*, Zhiyong Wu, Xipeng Qiut, Lingpeng Kong",
"Shark-NLP, Shanghai AI Laboratory Fudan University National University of Singapore",
"... | {
"arxiv_id": "2310.00280",
"language": "zh",
"page_idx": 0
} | ||
2310.00280-zh-1 | 2310.00280-zh-1.jpg | ${\mathrm{~push0~g~0~G}}\underline{{\mathrm{Preprint}}}$
$\therefore$pop
0.0.5Zheng et al.ˋpop, 'pushrgb 00.0.52023à pop)也被用来引出更准确的预测。然而，这些方法被限制在一个静态黑盒$\mathbb{E}($ pushrgb 0.0.5Yao et al.'pop, 'pushrgb 0.0.52023b'pop)中，其中 LLM 仅依赖于其内部表示来生成响应，并且容易生成不可靠的答案$\overline{\mathrm{E}(\text{pushrgb}00.5}$Ji et al. 'pop, 'pushr... | [
[
0.1748046875,
0.037261962890625,
0.313720703125,
0.0309295654296875,
0.314453125,
0.048614501953125,
0.176513671875,
0.054931640625
],
[
0.904296875,
0.0404052734375,
0.93896484375,
0.0404052734375,
0.93896484375,
0.05303955078125,
0.904296875,
0... | [
"pushO g O GPreprint",
"pop",
"00.5Zhengetal.pop，pushrgb000.52023apop）也被用来引出更准确的预测。然而，这些方",
"法被限制在一个静态黑盒E(pushrgb000.5Yaoetal.pop，pushrgb000.52023bpop）中，其",
"中LLM仅依赖于其内部表示来生成响应，并且容易生成不可靠的答案ElCpushrgb000.5Ji",
"et al. pop, pushrgb 0 0 0.52023 pop; pushrgb 0 0 0.5Yin et al. pop,pushrgb 0 0 0.52023b'pop)",
... | {
"arxiv_id": "2310.00280",
"language": "zh",
"page_idx": 1
} | ||
2310.00280-zh-2 | 2310.00280-zh-2.jpg | ${\mathrm{~push0~g~0~G}}\underline{{\mathrm{Preprint}}}$
$\therefore$pop
性和忠诚度，'pushrgb 0 0 0.5He et al.'pop( pushrgb 0 0 0.52022 pop) 和'pushrgb 0 0 0.5Wang et al. pop( pushrgb 0 0 0.52023c'pop) 利用外部知识库。最近，'pushrgb 0.0.5Gao et al.'pop ( pushrgô $\hat{0}$ 0 0.52023 pop) 通过提供引用的段落来确保生成文本的事实正确性和可验证性
另一行是将推理任务委托给外部工具$\op... | [
[
0.1748046875,
0.037261962890625,
0.31201171875,
0.032196044921875,
0.313720703125,
0.04986572265625,
0.176513671875,
0.054931640625
],
[
0.904296875,
0.037261962890625,
0.939453125,
0.0404052734375,
0.93701171875,
0.054931640625,
0.90185546875,
0... | [
"pushO g OGPreprint",
"pop",
"性和忠诚度，pushrgb000.5Heetal.pop（pushrgb000.52022pop)和pushrgb000.5Wang",
"etal.pop（pushrgb000.52023cpop）利用外部知识库。最近，pushrgb000.5Gaoetal.pop",
"（pushrgb000.52023pop）通过提供引用的段落来确保生成文本的事实正确性和可验证性",
"另一行是将推理任务委托给外部工具FCpushrgb000.5Oinetal.poppushrgb00",
"0.52023°pop），这些工具通常用于解决数值问题。其中... | {
"arxiv_id": "2310.00280",
"language": "zh",
"page_idx": 2
} | ||
2310.00280-zh-3 | 2310.00280-zh-3.jpg | push0 g 0 G$\underline {\mathrm{Preprint}}$
$\therefore$pop
和预测的质量。结论由$h(c_{\mathrm{blue}}^t,p_{\mathrm{blue}}^t,c_{\mathrm{green}}^t,p_{\mathrm{green}}^t)$在所有回合中确定，确保全面评估和更明智
的最终决定
与以前的作品不同$\mathbb{E}($ pushrgb 00.5Du et al.'pop, pushrgb 000.52023 pop; 'pushrgb 000.5Liang et al.'pop, 'pushrgb 0000.52023 pop; 'push... | [
[
0.1748046875,
0.037261962890625,
0.313720703125,
0.0309295654296875,
0.314453125,
0.048614501953125,
0.176513671875,
0.054931640625
],
[
0.904296875,
0.037261962890625,
0.94140625,
0.0404052734375,
0.93896484375,
0.054931640625,
0.90283203125,
0.... | [
"pushO g O GPreprint",
"pop",
"和预测的质量。结论由h(CbluePblue，Cgreen，Pgreen）在所有回合中确定，确保全面评估和更明智",
"的最终决定",
"与以前的作品不同pushrgb000.5Duetal.pop",
"pushrgb 0 0 0.52023pop;pushrgb 0 0 0.5Liang",
"et al.pop, pushrgb 0 0 0.52023pop; pushrgb 0 0",
"0.5Xiong et al. pop,pushrgb 0 0 0.52023 pop)在",
"LLM之间采用“辩论”，我们的方法采用小... | {
"arxiv_id": "2310.00280",
"language": "zh",
"page_idx": 3
} | ||
2310.00280-zh-4 | 2310.00280-zh-4.jpg | push0 g 0 G$\underline {\mathrm{Preprint}}$
-pop
## 3.3 RETRIEVE
3.3 Retrieve
在最后的工作线程中，我们深人研究了检索模式，以通过协作确定最忠实的答案。虽然以前基于多数投票机制$\mathbb{E}($pushrgb 00.5Wang et al.'pop, 'pushrgb 000.52023e' pop; ' pushrgb 000.5Fu et al. pop, 'pushrgb 000.52023c'pop) 的策略可以缓解诸如波束搜索$\textcircled{\mathrm{E}}($ pushrgb 000.5Li & Jurafsky... | [
[
0.1748046875,
0.037261962890625,
0.313720703125,
0.0309295654296875,
0.314453125,
0.048614501953125,
0.176513671875,
0.054931640625
],
[
0.90283203125,
0.037261962890625,
0.939453125,
0.039154052734375,
0.93896484375,
0.054931640625,
0.90185546875,
... | [
"pushO g 0 GPreprint",
"pop",
"3.3",
"",
"RETRIEVE",
"3.3Retrieve",
"在最后的工作线程中，我们深人研究了检索模式，以",
"通过协作确定最忠实的答案。虽然以前基于多数投",
"票机制E(pushrgb000.5Wangetal.pop,pushrgb00",
"0.52023e'pop; pushrgb 0 0 0.5Fu et al. pop, pushrgb 0 0",
"0.52023cpop）的策略可以缓解诸如波束搜索ECpushrgb",
"000.5Li& Jurafskypop,pushrgb 000... | {
"arxiv_id": "2310.00280",
"language": "zh",
"page_idx": 4
} | ||
2203.08383-zh-2 | 2203.08383-zh-2.jpg | 法，如提示(Liu et al.,2021),使 PLM 的参数保持不变，对于我们获取知识的目的是可取的。但是，我们发现，无论是进行微调还是提示一次生成一系列知识，PLM 都倾向于在此过程中失去其“思想链”,从而产生无关的事实或遭受幻觉
受多步推理问题的迭代性质的启发，我们在本文中探索了一种迭代提示框架，该框架针对给定的推理任务逐步从 PLM 中获取知识。我们在迭代提示中有两个需求：(1)在不同的推理步骤中，提示需要关注复杂查询的不同组件。(2)提示应将先前步骤中收集的知识适当地集成到当前步骤中；例如，在图 1 中示例的第 2 步期间，提示需要将实体“DavidLloydGeorge”(来自第1步中回顾的知识)与查询中未解决的部分“... | [
[
0.12005615234375,
0.0885009765625,
0.48779296875,
0.0885009765625,
0.48779296875,
0.09857177734375,
0.12005615234375,
0.09857177734375
],
[
0.50732421875,
0.08489990234375,
0.6767578125,
0.0872802734375,
0.6767578125,
0.1015625,
0.50732421875,
0.... | [
"法，如提示（Liuetal..2021），使PLM的参数保",
"2.1Problem Setup",
"",
"持不变，对于我们获取知识的目的是可取的。",
"给定一个复杂的查询9，我们的目标是驱动",
"但是，我们发现，无论是进行微调还是提示一",
"PLMM调用一系列简单的知识语句Cg",
"次生成一系列知识，PLM都倾向于在此过程",
"-",
"[c1，……,Cna]，足以决定对q的响应。特别是，我",
"中失去其“思想链”，从而产生无关的事实或",
"们专注于开发提示方法，其中M的参数是固",
"遭受幻觉",
"定的，我们的目标是构造提示T，引导M调",
... | {
"arxiv_id": "2203.08383",
"language": "zh",
"page_idx": 2
} | ||
2203.08383-zh-3 | 2203.08383-zh-3.jpg | <FigureHere>
Figure 2: Our context-aware prompter design. The prompter contextualizes a set of special tokens w.r.t. the current step context $q,c_1,...,c_{j-1}$ to get the resulting prompts, which steers the PLM to recall the next piece of knowledge $c_j$ .
示)。提示被设计为一组特殊令牌的上下文(由 Prompter)。r。t。当前步骤输入上下文，通过可训练矩阵线性投影到 ... | [
[
0.515625,
0.0885009765625,
0.87890625,
0.0885009765625,
0.87890625,
0.09857177734375,
0.515625,
0.09857177734375
],
[
0.51123046875,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.51123046875,
0.11639404296875
]... | [
"中的给定示例所示。因此，我们遵循先前的工",
"作 (Zhong et al., 2021; Li and Liang, 2021; Lester",
"PLMDecoderBlock (fixed",
"etal.2021QinandEisner，2021），并将提示建",
"模为连续的虚拟令牌，而不是离散的令牌",
"PLMEncoderBlock(fixed",
"",
"2.4Learning and Inference",
"",
"ec",
"rom",
"我们使用老师强迫进行模型训练，即，在每个",
"步骤中，将该步骤的地面实况上下文（查询和",
"以前的知... | {
"arxiv_id": "2203.08383",
"language": "zh",
"page_idx": 3
} | ||
2203.08383-zh-4 | 2203.08383-zh-4.jpg | 我们的最终训练集，这在很大程度上减少了测
试/训练重叠(更多细节在 §4.2)
多跳 QA 数据集，其中包含带注释的证据路径。数据集包含4。6k 示例(2。4k列车，2。2k开发)构建在 HotpotQA 之上，作者利用众包的努力，以简单的主语-动词-宾语自然语言句子的形式收集证据路径。同样，我们将开发集(没有给出测试集)随机拆分为我们的开发集和测试集(1。每个 1k 个样品)。我们将问题用作查询$q$,并将带注释的证据句子用作$C_q$
LoT (Talmor et al.,2020b)。数据集涉及从 ConceptNet 和 WordNet 构建的一组分类关系的推理。每个示例都包含一个假设(e。g。,“鲸鱼有一个肚脐”),我们将... | [
[
0.118408203125,
0.0872802734375,
0.48779296875,
0.0872802734375,
0.48779296875,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.497802734375,
0.0914306640625,
0.50537109375,
0.0914306640625,
0.50537109375,
0.09442138671875,
0.497802734375,
... | [
"我们的最终训练集，这在很大程度上减少了测",
"",
"PLM的一项技能如先前的工作所示，。我们",
"试/训练重叠（更多细节在$4.2）",
"还在向读者提供地面实况上下文（Oracle-RD）",
"R4C（Inoueetal.，2020）。R4C是最近的另一个",
"时报告最终推理结果，作为外部评估的上限。",
"多跳QA数据集，其中包含带注释的证据路径。",
"不包括特定于关系的提示方法，如（Shinetal.",
"数据集包含4。6k示例（2。4k列车，2。2k开",
"2020; Zhong et al., 2021; Petroni et al., 2019),因",
... | {
"arxiv_id": "2203.08383",
"language": "zh",
"page_idx": 4
} | ||
2303.04791-zh-4 | 2303.04791-zh-4.jpg | Title Suppressed Due to Excessive Size
时互易晶格的连续极限。由于我们只能用一组有限的数字来表示$s(\cdot)$和$\hat{\Phi}(\|\cdot\|)$,因此我们需要自己施加一个合适的离散化方案。在我们的实现中，我们通过立方体素网格平铺截止区域，并通过每个体素内的平均值集来替换所有连续函数。离散化分辨率由体素旁长$\Delta$控制，这是一个与非周期性情况相关的新的超参数。在我们的主要实验中，我们对每个测试模型使用等分辨率$\Delta=0.2\hat{\mathrm{A}}^{-1}$。App. A.3 中报道了其他分辨率的消融研究。请注意，仅当体素网格位于随输入结构旋转的坐标系... | [
[
0.358642578125,
0.058074951171875,
0.615234375,
0.058074951171875,
0.615234375,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.09234619140625,
0.08837890625,
0.47216796875,
0.08837890625,
0.47216796875,
0.0985107421875,
0.09234619140625,
... | [
"Title Suppressed Due to Excessive Size",
"时互易晶格的连续极限。由于我们只能用一组有限的",
"Computationalcomplexity：我们的远程消息传递方",
"数字来表示s(）和d(II-），因此我们需要自已施加",
"一",
"案的三部分过程，在Fig.3中进行了说明，其计算复",
"个合适的离散化方案。在我们的实现中，我们通过立",
"杂度为ONatNk），其中Nat是结构中的原子数（或",
"方体素网格平铺截止区域，并通过每个体素内的平均",
"在PBC的情况下是超级单元），Nk是包含在远程总和",
"值集来替换所有连续函数... | {
"arxiv_id": "2303.04791",
"language": "zh",
"page_idx": 4
} | ||
2205.02959-zh-1 | 2205.02959-zh-1.jpg | # 来自次优演示的团队策略的半监督模仿学习
Sangwon Seo and Vaibhav V. Unhelkar
Rice University
$\{{\mathrm{~sangwon.seo,~vaibhav.unhelkar~}}\}{\mathrm{~@rice.edu}}$
# Abstract
We present Bayesian Team Imitation Learner $({\mathrm{BTIL}}),{\mathrm{an~imitation~learning~algorithm~to}}$ model the behavior of teams performing sequential tas... | [
[
0.374267578125,
0.11553955078125,
0.625,
0.11553955078125,
0.625,
0.1300048828125,
0.374267578125,
0.1300048828125
],
[
0.406005859375,
0.1331787109375,
0.59228515625,
0.1331787109375,
0.59228515625,
0.15087890625,
0.406005859375,
0.15087890625
... | [
"来自次优演示的团队策略",
"的半监督模仿学习",
"Sangwon Seo and Vaibhav V. Unhelkar",
"Rice University",
"[ sangwon.seo, vaibhav.unhelkar @rice.edu",
"无需进行不安全的探索。通过提供新颖的多代理模仿学",
"Abstract",
"习技术，这些技术受到现实世界的团队合作考虑因素的",
"启发，这项工作旨在使代理能够在顺序任务中建模，评",
"We present Bayesian Team Imitation Learner",
"估和改进人与人之间的团队合作。从数学上讲，模... | {
"arxiv_id": "2205.02959",
"language": "zh",
"page_idx": 1
} | ||
2205.02959-zh-2 | 2205.02959-zh-2.jpg | <FigureHere>
Figure 1: Movers domain. The human-robot team needs to $\bar{\text{move the boxes to the flag. Each box can be moved if and}}$ only if both agents pick it up and move it along the same $\operatorname{direction}.$
1991; Song et al., 2018] 对我们的解决方案进行基准测试。在我们的评估中，我们强调了收集团队演示的挑战， 方法是收集一个新的人-代理团队合作数据集，用于(a) 心... | [
[
0.51708984375,
0.06817626953125,
0.9111328125,
0.0694580078125,
0.9111328125,
0.08331298828125,
0.51708984375,
0.08209228515625
],
[
0.5146484375,
0.08209228515625,
0.91015625,
0.08331298828125,
0.91015625,
0.09783935546875,
0.5146484375,
0.09661... | [
"Multi-agentImitationLearning（MAIL)：尽管存在多",
"种MAIL算法，但现有技术中考虑的问题设置与本文考",
"虑的问题设置不同。从Sec.1开始，先前的MAIL技",
"术要么学习与多代理系统[Songetal.，2018;Linetal.",
"2019]的博奔论均衡相对应的行为，要么假设代理能力",
"[Bhattacharyyaetal.，2019]具有同质性，要么不考虑潜",
"在的性能塑造因素（例如心理模型或认知状态）。对潜",
"在状态进行建模的方法假设潜在状态在成员[Wang et",
"al.,2021;Ivanovicetal.,2018... | {
"arxiv_id": "2205.02959",
"language": "zh",
"page_idx": 2
} | ||
2205.02959-zh-3 | 2205.02959-zh-3.jpg | $\bullet$ $a_i\in A_i$ , is the set of actions $a_i$ available to the $i$-th
agent;
$\bullet$ $A=$ $\times _{i}$ $A_{i}$ is the set of joint actions, where
$a=[a_1,\cdots,a_n]$ denotes the joint action;
$\bullet$ $T_{s}( s^{\prime }| s, a) : S\times A\times S\to [ 0, 1]$denotes the state transi-
tion probabilities... | [
[
0.0931396484375,
0.07135009765625,
0.484375,
0.07135009765625,
0.484375,
0.08209228515625,
0.0931396484375,
0.08209228515625
],
[
0.53369140625,
0.0706787109375,
0.91259765625,
0.0706787109375,
0.91259765625,
0.083984375,
0.53369140625,
0.0839843... | [
"aEA;istheset of actionsa,availabletothei-th",
"member's decision a conditioned on their decision",
"agent;",
"factors (s,ri)",
": A= Xi A, is the set of joint actions,",
"where",
"参考运行示例，Alice和Rob的行为不仅取决于任务上",
"a=-[ai,.., an] denotes the joint action;",
"下文（MMDP状态），还取决于他们对下一个目标位置的",
". T,(s'|s, a... | {
"arxiv_id": "2205.02959",
"language": "zh",
"page_idx": 3
} | ||
1904.10638-zh-2 | 1904.10638-zh-2.jpg | 它实现了更好的结果，特别是对于大的重定向角度
# 3. Gaze Adaptation approach
我们用于用户特定注视自适应的总体方法如图。1 。它包括使用标记的训练样本微调通用神经网络。然而， 我们建议使用视线重定向模型生成额外的样本，而不是只使用很少 (少于 10) 的参考样本。由于此重定向模型是我们方法的主要组成部分，因此我们在$\color{red}{\mathrm{E3.1~}}$到E 3.4部分中对其进行了更详细的描述。然后在$\color{red}{3.5}$节中描述注视自适应部分
## 3.1. Gaze Redirection Overview
我们的凝视重定向框架如图所示。2。它由重定向网络本身和... | [
[
0.08416748046875,
0.0947265625,
0.439453125,
0.0947265625,
0.439453125,
0.10479736328125,
0.08416748046875,
0.10479736328125
],
[
0.50341796875,
0.091552734375,
0.8896484375,
0.0921630859375,
0.8896484375,
0.106689453125,
0.50341796875,
0.1060791... | [
"它实现了更好的结果，特别是对于大的重定向角度",
"<NEXMATHWELIN5。与直接合成相比，该策略将输",
"入的像素投影到输出，这保证了输人和输出将共享相",
"3. Gaze Adaptation approach",
"似的颜色和照明分布",
"我们用于用户特定注视自适应的总体方法如图。1。",
"对于训练，我们使用L1损耗来测量重定向输出I△g",
"它包括使用标记的训练样本微调通用神经网络。然而，",
"和地面实况GI之间的差异。因此，以间接监督的方式",
"我们建议使用视线重定向模型生成额外的样本，而不",
"学习生成重定向所需的逆翘曲场",
"是只使用很少（少... | {
"arxiv_id": "1904.10638",
"language": "zh",
"page_idx": 2
} | ||
1904.10638-zh-4 | 1904.10638-zh-4.jpg | 说足够真实。请注意，在本文中，我们只针对单眼图像凝视状态(以及重定向和自适应),而将全脸情况作为未来的工作
# 4.1. Experimental Setting
Datasets. 我们使用 ColumbiaGaze 数据集$\overline{\mathbb{E}}[28]$ 和MPIIGaze 数据集$\mathbb{E}[43]$ 进行实验。前者包含 56 人的注视样本，而后者包含 15 人的眼睛图像
Generic gaze estimator.作为我们的注视估计器，我们使用 GazeNet$\mathbb{E}[45]$ 。它基于 vgg16 架构。为了训练它，我们遵循 ColumbiaGaze 和 MPIIGaze... | [
[
0.08251953125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.08251953125,
0.1060791015625
],
[
0.50244140625,
0.08709716796875,
0.88623046875,
0.08837890625,
0.88623046875,
0.10162353515625,
0.50244140625,
0... | [
"说足够真实。请注意，在本文中，我们只针对单眼图像",
"Table 1.ColumbiaGaze dataset:gaze adaptation perfor",
"凝视状态（以及重定向和自适应），而将全脸情况作为",
"manc",
"error",
"approad",
"未来的工作",
"Cross",
"LinAdapSVRAdaFTAdap DiffNetRedFTAdap",
"#sample",
"Subjec",
"5.53",
"4.64",
"3.92",
"1",
"4.1.Experimental Setting",
"59",
"... | {
"arxiv_id": "1904.10638",
"language": "zh",
"page_idx": 4
} | ||
2203.15793-zh-0 | 2203.15793-zh-0.jpg | Instance Relation Graph Guided
Source-Free Domain Adaptive Object Detection
Vibashan VS, Poojan Oza, and Vishal M. Patel Johns Hopkins University, Baltimore, MD, USA
$\{$ vvishnu2,poza2,vpatel$36\}$ @jhu.edu
Abstract
Unsupervised Domain Adaptation (UDA) is an effective approach to tackle the issue of domain shift. S... | [
[
0.32177734375,
0.1331787109375,
0.64892578125,
0.1337890625,
0.64892578125,
0.1483154296875,
0.32177734375,
0.147705078125
],
[
0.24267578125,
0.1546630859375,
0.72607421875,
0.1558837890625,
0.72607421875,
0.173583984375,
0.24267578125,
0.172363... | [
"Instance Relation Graph Guided",
"Source-Free Domain Adaptive Object Detection",
"Vibashan VS. Poojan Oza, and Vishal M. Patel",
"Johns Hopkins University, Baltimore, MD, USA",
"[ vvishnu2,poza2,vpatel36 @jhu.edu",
"Abstract",
"Unsupervised Domain Adaptation (UDA) is an ef-",
"fective approach to tac... | {
"arxiv_id": "2203.15793",
"language": "zh",
"page_idx": 0
} | ||
2203.15793-zh-2 | 2203.15793-zh-2.jpg | 标签训练和信息最大化进行对齐。继分段任务 Liu et al.[48]之后，提出了一种基于自监督和知识转移的目标域自适应策略。对于目标检测任务，[41]提出了一种伪标签自训练策略，[28] 提出了通过以前的模型方法进行自监督特征表示学习
对比学习。无监督特征学习的巨大成功是由于对比学习，这归因于许多无监督任务[5,28,54]的巨大改进。对比学习通常通过最大化正对之间的一致性并最小化与负对的一致性来学习判别式特征嵌人。在[5,18,54] 中。在图像的批次中，锚图像经历不同的增强，并且该锚的这些增强形成正对和负对从给定批次中的其他图像中采样。后来，在利用特定于任务的语义信息的[32] 中，类内特征嵌人被拉在一起，并被排斥在跨类特征嵌... | [
[
0.08251953125,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.50341796875,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
... | [
"标签训练和信息最大化进行对齐。继分段任务Liuet",
"训练模型生成的伪标签在未标记的目标数据上更新模",
"al.【48]】之后，提出了一种基于自监督和知识转移的目",
"型。通过置信度阈值过滤伪标签，并使用可靠的伪标签",
"标域自适应策略。对于目标检测任务，41提出了一种",
"来监督检测器训练31]。更正式地说，对象检测模型",
"伪标签自训练策略，281提出了通过以前的模型方法进",
"的伪标签监督损失可以给出为：",
"行自监督特征表示学习",
"CsL=Cepn(at,yt)+ren(ct,t)",
"对比学习。无监督特征学习的巨大成功是由于对比学",
"习，这归... | {
"arxiv_id": "2203.15793",
"language": "zh",
"page_idx": 2
} | ||
2203.15793-zh-4 | 2203.15793-zh-4.jpg | 异表示为 GDL, 可以正式写成：
$\mathcal{L}_{GDL}=$KL$(\sigma(Z_{st}),\sigma(\tilde{Z}_{st}))$
+KL$(\sigma(Z_{te}),\sigma(\tilde{Z}_{te}))+$KL$(\sigma(Z_{st}),\sigma(Z_{te}))$,
其中 KL 表示 Kullback-Leibler 散度，$\sigma$表示 softmax 运算符。因此，最小化$\mathcal{L}_GDL$会监督 IRG 网络，进而学习实例关系矩阵({$\mathcal{E} )$
3.2.2 Graph Contrastive Loss (GCL)
实例成对... | [
[
0.5048828125,
0.09027099609375,
0.57275390625,
0.09027099609375,
0.57275390625,
0.10162353515625,
0.5048828125,
0.10162353515625
],
[
0.57275390625,
0.091552734375,
0.88134765625,
0.091552734375,
0.88134765625,
0.10162353515625,
0.57275390625,
0.... | [
"Table 1.",
"Quantitativeresults(mAP)for Cityscapes-",
"FoggyCityscapes. S: Source only, O: Oracle, UDA: Unsu-",
"异表示为GDL，可以正式写成",
"LGDL = KL(α(Zst),o(Zst))",
"pervised Domain Adaptation, SFDA: Source-Free Domain",
"+ KL(α(Zte), α(Zte)) + KL(α(Zst), α(Zte));",
"Adaptation.",
"其中KL表示Kullback-Leibler散... | {
"arxiv_id": "2203.15793",
"language": "zh",
"page_idx": 4
} | ||
2212.03019-zh-0 | 2212.03019-zh-0.jpg | Style transfer and classification in hebrew news items
Nir Weingarten
Reichman university, Israel nir. weingarten@ post. idc. ac. il
# 1 Introduction
Tsarfaty 等人将希伯来语归类为形态丰富语言(MRL)。(2010)。因此，前缀和后缀会附加到单词上以更改其语法含义。Tsarfaty 等人 (2019) 已经证明了这种结构会导致语言中固有的形态歧义。因此，除其他外，希伯来语 NLP 算法落后于其他非 MRL 语言
NLP 的最新发展以及希伯来语 NLP 领域的新研究带来了广... | [
[
0.15283203125,
0.1767578125,
0.7548828125,
0.177978515625,
0.7548828125,
0.1956787109375,
0.15283203125,
0.1944580078125
],
[
0.38818359375,
0.2215576171875,
0.51708984375,
0.2215576171875,
0.51708984375,
0.236083984375,
0.38818359375,
0.23608398... | [
"Style transfer and classification in hebrew news items",
"Nir Weingarten",
"Reichman university, Israel",
"nir.weingarten@post.idc.ac.il",
"1Introduction",
"Tsarfaty等人将希伯来语归类为形态丰富语言（MRL）。（2010)。因此，前缀和后缀会附",
"加到单词上以更改其语法含义。Tsarfaty等人（2019）已经证明了这种结构会导致语言中",
"固有的形态歧义。因此，除其他外，希伯来语NLP算法落后于其他非MRL语言",
"NL... | {
"arxiv_id": "2212.03019",
"language": "zh",
"page_idx": 0
} | ||
2212.03019-zh-1 | 2212.03019-zh-1.jpg | 续计算每个输入$x$的查询向量与序列中所有其他输入的键向量的点积。这产生了一个新的向量$A$输入序列的长度。$A$用作权重向量，以与交叉注意相同的方式计算上下文向量$M$ 。因此，转换器可以轻松快速地从序列的任何部分维护上下文。堆叠许多注意力可以产生更强大的学习，并且该架构非常适合迁移学习
最近在 NLP 中非常流行的变压器改编是 BERT-用于语言理解的双向变压器 (Devlin, Chang, Lee, Toutanova2018)。在这项工作中，BERT 既用于分类器，也部分用于文本生成器。BERT 是一种体系结构，旨在允许在未标记的文本上双向预训练强大而健壮的语言模型 (即在所有层中对左右上下文进行条件调节)。这种预先训练的... | [
[
0.136474609375,
0.121826171875,
0.76708984375,
0.12249755859375,
0.76708984375,
0.1363525390625,
0.136474609375,
0.1351318359375
],
[
0.1380615234375,
0.141357421875,
0.7685546875,
0.141357421875,
0.7685546875,
0.1546630859375,
0.1380615234375,
0... | [
"续计算每个输入的查询向量与序列中所有其他输入的键向量的点积。这产生了一个新",
"的向量A输入序列的长度。A用作权重向量，以与交叉注意相同的方式计算上下文向量",
"M。因此，转换器可以轻松快速地从序列的任何部分维护上下文。堆叠许多注意力可以",
"产生更强大的学习，并且该架构非常适合迁移学习",
"最近在NLP中非常流行的变压器改编是BERT-用于语言理解的双向变压器（Devlin，",
"Chang，Lee，Toutanova2018）。在这项工作中，BERT既用于分类器，也部分用于文本",
"生成器。BERT是一种体系结构，旨在允许在未标记的文本上双向预训练强大而健壮的",
"语言模型（即... | {
"arxiv_id": "2212.03019",
"language": "zh",
"page_idx": 1
} | ||
2212.03019-zh-2 | 2212.03019-zh-2.jpg | (数据点)之间构建加权图，其中权重是两个点连接的可能性(优化权重)。这是在全局和局部考虑之间取得平衡的情况下完成的。在本地计算权重，以便一个点“靠近”其 n 个最近的邻居 (在人口稀少的地区 , 本地距离的权重较小 ,而在人口稠密的地区 , 本地距离的权重更大)。在全球范围内，随着距离的增加，它降低了连接的可能性。2。构造一个低维(布局)图，尽可能类似于加权，优化，图
# 3 Solution
3.1 General approach
3.1.1 Data collection and preprocessing
使用专门构建的 pythonWeb 刮刀从在线新闻门户档案中收集数据。我们决定抓取新闻数据，而不是使用现有的希伯... | [
[
0.1380615234375,
0.12249755859375,
0.76708984375,
0.12249755859375,
0.76708984375,
0.1363525390625,
0.1380615234375,
0.1363525390625
],
[
0.1396484375,
0.141357421875,
0.7685546875,
0.141357421875,
0.7685546875,
0.1546630859375,
0.1396484375,
0.1... | [
"（数据点）之间构建加权图，其中权重是两个点连接的可能性（优化权重）。这是在全局和",
"局部考虑之间取得平衡的情况下完成的。在本地计算权重，以便一个点“靠近”其n个",
"最近的邻居（在人口稀少的地区，本地距离的权重较小，而在人口稠密的地区，本地距离",
"的权重更天）。在全球范围内，随看距离的增加，它降低厂连接的可能性。2。构造一个低",
"维（布局）图，尽可能类似于加权，优化，图",
"3",
"Solution",
"3.1",
"General approach",
"3.1.1",
"Data collection and preprocessing",
"使用专门构建... | {
"arxiv_id": "2212.03019",
"language": "zh",
"page_idx": 2
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.