3 algorithm

3.1 Different from previous attention based MIL algorithms, the method embeds two kinds of attention mechanisms into one network for MIL tasks.
In our method, the attention and self-attention mechanisms are employed because they can be combined naturally.
这两句话意思重复、啰嗦，前后都在说两种注意力机制
-》
Different from previous attention based MIL algorithms, ASMI embeds the attention and self-attention mechanisms into one network.

3.2 The attention block can assign weights for each instance in the bag and aggregate each bag into a single vector.
can可以不用，陈述事实，不需要可以、或许等委婉的词
weights vs each instance
bag 读起来有些冗余
-》
The attention block assigns a weight to each instance and aggregates each bag into a single vector.

3.3 The self-attention block is constructed for qualify the influence of each instance to the fused single vector.
对照上一句话，用动词来说明直接描述即可
度量：measure
-》The self-attention block measures the influence of each instance to its fused single vector.

3.4 With the information of each instance, the distinguishability of vectors generated by self-attention block can be increased.
这个段落是概述算法的过程，简单描述两个部分做什么就行，这句话描述效果，放在这里还不是时候，而且没必要让这一段很长。

3.5 A ： DATA Model 这里， yi \in {0, …, K-1}
B ： where N represents the cardinality. b is partitioned according to respective labels into { b0; b1; …; bL-1 }
L和K不统一

3.6 where bi represents the mean vector of bi and d*(bi) represents the mean distance between each vector in bi and d(; ) represents the Euclidean distance.
三个and，前面一个and换成逗号就可以了
-》
where bi represents the mean vector of bi，d*(bi) represents the mean distance between each vector in bi and d(; ) represents the Euclidean distance.

3.7 The denominator in Eq. (1) simply quantifies the separation degree between each vector set belonging to different labels.
simply不需要了，我们是在说自己的好，没必要拉自己的台
quantify -》 measure
-》
The denominator in Eq. (1) measures the separation degree between each vector set belonging to different labels.

3.8 discrimination
你前后都是用的distinguishability

3.9 Notably, it is uncertain that there is a strictly positive correlation between the distinguishability and the classification accuracy.
But the higher distinguishability usually followed by higher classification accuracy.
这两句话有排斥关系，需要斟酌
闵老师指导：这种问题可论述，不可证明。
类似于SVM 最大化间距，泛化性更好（测试集上效果更好）。
distinguishability -> margin -> generalizability -> more accurate in unknown instances
Many machine learning schemes such as SVM obtain better generalization ability through maximizing the classification margin.
Similarly, here we try to maxmize distinguishability for the same purpose.

3.10 The main task of the attention block is to quantify the contribution of each instance to the bag Bi and obtain a
fused vector hi.
问题1：each instance 有歧义
->
The main task of the attention block is to measure the contribution of each instance \mathbf{x}_ij to the bag Bi and obtain a fused vector hi.

问题2，fig. 1 和 2对照，因fig. 1简化成框架了，那么fig 2应交代清楚，让读者能对应上fig2 是fig1的哪个部分
以C段落目前的描述：
1）fig2就是attention block，那在fig1中看来，attention block的输入是包Bi，但是Fig2是实例。
2）fig1中attention block的输出是hi，但Fig2是系数
因此，C这里还需要再优化，并且建议fig1那里，还是要把两个注意力块的输入输出标注清楚，并与fig2和fig3照应

问题3：
根据问题2的分析，具体举例说明，fig2所示的attention block只能量化实例对包的贡献，不能获得hi
而E段，这句话：And the fused vector hi can be obtained based on the learned weights. 应该才是说准确了的
闵
-》
The main task of the attention block is to measure the contribution of each instance xij to the bag Bi.
Then the fused vector hi can be obtained based on the learned weights.

3.11 The dimension of the new feature space depends on the richness of information contained in the original features.
the richness of information这个说得不具体
闵： demension是量化的东西而richnes是一个形容词，没法支撑demension
如果是下面的意思，则应该放到实验部分去说
-》
For sparser data, the dimension of the new feature space should be set to lower.

3.12 The illustration of calculation process for one attention weight.
-》
注意名词的词性可数名词要有冠词数量词等，要么就是负数同理检查fig. 3下面
A illustration of the calculation process for one attention weight.

3.13 P( ) is firstly employed to transform…
Q() is utilized …
the softmax function is used to …
文中基本都用的这用句式，被动是一种静态关系，可以多采用动态关系描述，用主动态如：P( ) transforms…

3.14 As the result, the fused vector hi of Bi can be computed as the sum of product between …:
建议通检查一遍 can …，描述算法的过程，主动态或者被动态，直接用就是了。当需要用婉转、可能等来委婉表达的时候，再用

3.15 文章通篇类似这样的表示：attention based MIL method
都应该加短划线把
attention-based MIL method

3.16 However, it is possible that a large weight is assigned to an instance x_ik, k \in [1; : : : ; ni] which is inconsistent with the bag
Bi.
-> it is possible … 是可以的，但如果直接用情态动词来缩短语句会更好。这里可以和3.14建议不用情态动词的情况做一个对比
前面表示包Bi中的实例，是用的x_ij吧，这里又多引入了一个k变量
->
However, a large weight might be assigned to an instance x_ik, k \in [1; : : : ; ni] which is inconsistent with the bag Bi.

3.17 To alleviate this problem, the main task of self-attention block is to use information of each instance to generate an enhanced fused vector bi, which has more representation power than hi.
长句。利用每个示例的信息产生bi，这个不准确， attention block也是利用实例的信息，所以没有体现出不同
你看下是不是下面的意思
->
To alleviate this problem, the self-attention block is proposed.
The enhanced fused vector bi is obtained by exploiting the contribution of each instance x_ij to the fused vector h_i.
This introduces two advantages. One… The other is …
Therefore, bi is more representative than hi.
-》
闵老师
To alleviate this problem, the self-attention block is proposed.这句话没有意义
-》
To address this issue, we design the self-attention block to obtain the enhanced fused vector bi by exploiting the contribution of xij to hi.

3.18 extractions -》 extractors

3.19 As mentioned above, there are two types of mechanisms that are respectively utilized in the proposed algorithm.
-》
As mentioned above, ASMI employs attention and self-attention mechanisms.

3.21 我认为E部分，第一段和第二段的前面部分，不是Analysis，更像是conclusion
从The former computes attention weights…开始，能勉强靠analysis这个方向。
闵：
"分析"一般都是时间和空间复杂度分，以式子为主干，而不是用文字本身分析
Analysis: time/space complexity analysis
分成几点。
如果没有充足的分析支撑，可以用 Discussions

论文写作——韬ASMI29 algorithm部分语句修改相关推荐

论文写作——韬ASMI29 introduction、related work两部分语句修改
文章目录 1 introduction 2 related work 1 introduction 1.1 Therefore, a bag is described by a single bag ...
论文写作的不同阶段，该如何修改？
首先我们要确定自己的稿件论文,是处于哪个写作时期,未开稿?初稿?修改稿?还是定稿了? 未开稿期间未开稿之前,会涉及大量文献论文的阅读,建议多看看国外的文献,说实话,可以大大降低重复率,如果是文献翻译 ...
Algorithm：数学建模大赛之数学建模基础(经验/技巧)、流程(模型准备/模型假设/建模/求解/分析/优化/预测/评价)、论文写作(意义/摘要/关键词/问题重述和模型假设/建模/文献)之详细攻略
Algorithm:数学建模大赛之数学建模基础(经验/技巧).流程(模型准备/模型假设/建模/求解/分析/优化/预测/评价).论文写作(意义/摘要/关键词/问题重述和模型假设/建模/求解/结论/参考文 ...
修改正文中参考文献标注_论文写作中怎样正确插入参考文献，引用文献如何标注？...
论文写作中怎样正确插入引文文献,引用文献如何标注?不管是大学毕业生还是期刊/评职称的我们在面对撰写论文时,参考文献的引用是必不可少的.参考文献的引用可以给论文增添很多的光彩.正确的在论文中引用参考问下 ...
还在修改博士论文？这份《博士论文写作技巧》为你指南
点上方蓝字计算机视觉联盟获取更多干货在右上方 ··· 设为星标 ★,与你不见不散仅作分享,不代表本公众号立场,侵权联系删除转载于:专知,网络 AI博士笔记系列推荐周志华<机器学习> ...
Algorithm：数学建模大赛(CUMCM/NPMCM)之数学建模(经验/技巧)、流程(模型准备/模型假设/建模/求解/分析/优化/预测/评价)、论文写作(意义/摘要/关键词/问题重述和模型假设/建
Algorithm:数学建模大赛(CUMCM/NPMCM)之数学建模(经验/技巧).流程(模型准备/模型假设/建模/求解/分析/优化/预测/评价).论文写作(意义/摘要/关键词/问题重述和模型假设/建 ...
论文写作——5 introduction与语言逻辑（以韬韬同学ASMI27版为例）
文章目录 1 introduction 的节奏安排 2 先通篇检查: 3 常见的要大写的地方 3 第一段的问题 4 第二段的问题 5 第三段的问题 6 第四段的问题 7 第五段问题 8 第六段问题今 ...
论文写作和修改的技巧
论文写作是目前本科生.研究生.博士生的一个必备工作,我第一次写一篇能够发表的论文经过很多次的修改,而不是一蹴而就的.这其中,我的指导老师给我了我很大的帮助. 他告诉我一篇论文写好以后,你不要着急投出去 ...
科研、论文写作及修改
科研.论文写作及修改详细内容请查阅下文英文版科研进行研究的不同方式: a) 创建一个研究领域(例如,拓扑绝缘体等) b) 创建一个主题(例如,无线通信中的 MIMO 等) c) 关注热门话题(例 ...

论文写作——韬ASMI29 algorithm部分语句修改

文章目录

3 algorithm

论文写作——韬ASMI29 algorithm部分语句修改相关推荐

最新文章

热门文章