论文写作——韬ASMI29 introduction、related work两部分语句修改

1 introduction

1.1 Therefore, a bag is described by a single bag label because the instances generally do not have labels
1） 准确来说，应该是标签未知
2） a bag is described by a single bag label 意思绕
-》
Therefore, only the bag has a label and the labels of its instances are usually unknown
1.2 Based on the connection between the bag and its instances, the instancebased paradigm [11], [12] predicts bag labels using instancelevel classifiers.
感觉instance-based 方法没有说到点子上，connection有点泛，还需要说具体

1.3 The attention mechanism effectively decreases the influence of negative instances in positive bags. Unfortunately,…
建议换成 can。 这是别人的工作，而且马上就要说不足，没必要加effectively

1.4 It is possible that the fused vectors transformed from positive bags will have more negative characteristic as a result of this.
1）as a result of this感觉有点怪，this指代的是Unfortunately,…那句话吗？
as a result of this不需要， 因为不用，意思也很明确，是上一句话带来的可能结果，还让句子变短
2）建议
-》 Therefore, the fused vectors transformed from the positive bags may have more negative characteristics.

1.5 1) In the attention block, a learnable linear transformation is first applied to the instances in a given bag. Then a single layer
network is employed to learn the attention weight of each instance. Accordingly, a fused vector is calculated as the sum of products between transformed instances and their corresponding weights.
3) With the embedded information, self-attention mechanism …
ym：
第1部分主要在说内部怎么做的，涉及到每一步的细节
第2部分也在说内部怎么做，相对略——放introduction，我觉得ok
第3部分主要在说引入后的好处，没有涉及how——跟前面不匹配

建议1，弱化第1部分的描述，都不说那么细；则第3部分也就算统说，后面再慢慢描述。
建议2，表达如下意思：
第1部分 着重表达：通过学习实例对包的贡献，获得实例的权重，得到了输出
第3部分 ：通过学习其他实例对融合向量的影响，优化权重，才有后面的好处

2 related work

2.1 As we have seen yet, BP-MIP [13] firstly employ BP algorithm for MIL, which performs back propagation on the instances with a maximum
loss function
As we have seen yet多余了，去掉； employs； 长句换成短句
-》BP-MIP [13] first employs BP algorithm for MIL.
It performs back propagation on the instances with a maximum loss function.

2.2 Furthermore, Convolutional Neural Networks (CNNs) based MIL algorithms MIL-FCN [21] and WSCN are proposed to select latent instances for learning semantic information from weak image-level labels.
MIL-FCN [21] and WSCN放前面，主语更突出些;
MIL algorithms可以不写了，句子已经较长了
are proposed to有点累赘，可以不写
-》
Furthermore, MIL-FCN [21] and WSCN based on Convolutional Neural Networks (CNNs) select latent instances for learning semantic information from weak image-level.

2.3 However, maximum based networks have one flaw:instances are simply labeled as the bag to which they belong.
意思似乎不对。标记为所属的包吗？
-》
However, the disadvantage of the maximum-based network is that the label of the bag is directly assigned to its instances.

2.4 To alleviate this issue, several alternative pooling functions are proposed and embedded into MIL networks, such as MILBoost [23], ISR [24], and LSE [25].
1）缓解问题，搭配不当 -》 to address this issue. 或 to alleviate this drawback.
to address this issue. 比较一般化，更强调自己的完整方案。
to alleviate this drawback. 比较针对性强，补某一个漏洞，偏小器。
2） and 前后没有并列关系，proposed也没有必要，这些算法本来就是提出来解决问题的，所以直接了当说方法怎么用就可以了
-》
To address this issue, several alternative pooling functions are embedded into MIL networks, such as MILBoost [23], ISR [24], and LSE [25].

2.5 Attention-based deep multiple instance learning [17] designs a two-layered neural network to allocate learnable weights to instances and uses the sigmoid function
to predict bag probability.
长句。
-》
Attention-based deep multiple instance learning [17] designs a two-layered neural network.
It allocates learnable weights to instances and uses the sigmoid function to predict bag probability.

2.6 In order to obtain better performance on bag prediction and instance interpretation, Loss-attention [26] associates the attention mechanism with the loss function for smoothing the training process and furthermore to simultaneously learn instance and bag predictions.
长句，36个单词了
-》
这句话本质上是上一个自然段：MIL with Attention的一个方法。建议并到上一个段，这样也不用写In order to…，
后面的however… 可以另起一段
-》
Loss-attention [26] associates the attention mechanism with the loss function.
The purpose is to smooth the training process and simultaneously learn instance and bag predictions.

2.7 As mentioned above, instances whose information is inconsistent with their bag may be assigned large weights.
-》
As mentioned above, a instance whose information is inconsistent with its bag may be assigned large weight.

2.8
Different from the previous methods, our proposed algorithm fuses two attention mechanisms to efficiently exploit the influence of each instance to its fused vector.
这两句话和前后句子的意思重复，而且长。你看下以下意思到位不
Different from the previous methods, our proposed algorithm fuses two attention mechanisms.
This can simultaneously exploit the contribution of each instance to its bag and fused vector.