《科研实验室-Rebuttal经验总结》由会员分享，可在线阅读，更多相关《科研实验室-Rebuttal经验总结（24页珍藏版）》请在金锄头文库上搜索。
1、Rebuttal 经验总结 目录 Review和Rebuttal介绍 1 Review中的问题整理 2 Rebuttal撰写与案例分析 3 Rebuttal的精简和打磨 4 目录 Review和Rebuttal介绍 1 Review中的问题整理 2 Rebuttal撰写与案例分析 3 Rebuttal的精简和打磨 4 Review和Rebuttal介绍 2 Rebuttal 在最终决定文章是否被接收前 作者可以在Rebuttal 中回答审稿人的疑问 Rebuttal写得好可以提升文章接收的可能性 1 审稿意见 Review 会议论文投稿之后一般3个月之内会出 Review 其中包含3 4个审稿
2、人的评审意见 Review和Rebuttal 0123456789 Score Number of submissions Oral Poster Workshop Reject Table 1 ICLR2018 Scores Decisions 还可以抢救 Review和Rebuttal介绍 Rebuttal的一般要求 1 字数限制 被严格限制在500 1000 words以内 2 时间限制 通常有5 8天的准备时间 3 模版使用 严格按照要求使用会议提供的模版 4 可以增加图 表 证明回答审稿人的问题 5 Rebuttal中不能增加新的contributions 6 确保Rebuttal匿
3、名 不暴露作者信息 7 不允许通过引入链接访问网站的形式添加额外材料 强度很高 可以补充 实验结果 目录 Review和Rebuttal介绍 1 Review中的问题整理 2 Rebuttal撰写与案例分析 3 Rebuttal的精简和打磨 4 Review中的问题整理 第一步 整理总结审稿人的正面和负面评价 正面评价的作用 1 他山之石可以攻玉 例如可以用来回复质疑novelty的问题 2 向chair展示文章的亮点 通常放在Rebuttal最前面 正面评价的例子 1 As far as I know the idea is novel and the development of the
4、idea is well motivated 2 Section 3 3 and 3 6 also provides simple intuitive but useful insights 3 The overall learning framework makes sense and it consists several interesting components 4 I think the idea of this paper is interesting and worth a bigger audience for sure Review中的问题整理 整理负面评价中的问题 Rev
5、iewer 1 1 However I don t agree that X is novel I ve seen it in Foo 2017 2 Intuitively it is a discontinuous operation so is it suitable for parallel implementation 3 Sharp smooth window sets are decided by the mean value of different channels whether it is reasonable or not 4 Inference time frame p
6、er second FPS and model size are also very important criterions which should be considered Paper rating Borderline Review中的问题整理 第二步 总结需要补的实验 第一时间跑实验 1 Inference time frame per second FPS and model size are also very important criterions which should be considered 2 A simple baseline that the authors
7、 could compare with is to learn a latent representation using VAE and then use its decoder as an initialization point for GANs 3 How sensitive would the method be to the number of nearest neighbors used for local coordinate coding 4 Inception score IS cannot evaluate the generalization ability as si
8、mply memorizing all data gives highest score 需要补充FPS和model size的对比结果 需要对比审稿人描述 的方法 需要对某个超参数进 行敏感度分析 IS不够好 需要用其 他指标评价方法性能 Review中的问题整理 第三步 归纳审稿人的共性问题 避免重复回答 Reviewer 1 Inference time frame per second FPS and model size are also very important criterions which should be considered Reviewer 2 It is sti
9、ll unknown whether the proposed method is more efficient than regular convolution in practical usage Reviewer 1 和 Reviewer 2 都在问模型的实际计算效率 可以合并回答 总结回顾 第一步 整理总结审稿人的正面和负面评价 第二步 总结需要补的实验 第一时间跑实验 第三步 归纳Reviewers的共性问题 避免重复回答 Review的问题整理思路 目录 Review和Rebuttal介绍 1 Review中的问题整理 2 Rebuttal撰写与案例分析 3 Rebuttal的精简
10、和打磨 4 Rebuttal 结构 We thank all reviewers for their valuable feedback and address their main concerns as follows 对reviewer表示 尊重和感谢 All reviewers recognize our novelty that we build a new kind of xxx method R1 and this paper is interesting and worth a bigger audience for sure R2 用reviewer的话向 chair展示文章
11、的亮点 和重要性 optional General Response Q1 Difference with xxx method Unlike xxx our method Q2 Results of xxx We apply our method to xxx and show the results in Table A 共性问题要形 成共性回答 Response to Reviewer 1 R1 Q1 Training complexity of xxx The complexity of xxx Q2 Advantage of xxx Our method provides an ef
12、fective way 回答每个reviewer的 问题 问题案例分析 问题案例分析 提问类问题 首先直接回答 yes or no 之后加一个简短的解释 案例 Q Intuitively it is a discontinuous operation so is it suitable for parallel implementation A CAC is suitable for parallel implementation We implement CAC by dividing the matrix multiplication in convolution into two ind
13、ependent matrix multiplications which can be computed in parallel to further improve the efficiency 回答yes or no简短的解释 问题案例分析 解释类问题 抓住审稿人真正疑惑的点进行解释 切忌空谈赘述 案例 Q In figure 3 the visualization suggests that higher level contains mostly sharp regions I cannot understand why it is the case 审稿人不理解为什么大 部分区域都
14、是sharp的 A Higher level features contain mostly richer information than shallower features thus a small ratio may incur performance drop for higher level layers In fact from Fig 3 the higher level feature map and gradients tend to be denser rather than sharper 实际上是denser 而不是sharper 问题案例分析 实验结果类问题 1 可
15、以得到的结果 直接展示 说明方法有效性 2 无法得到的结果 简单说明原因 案例 可以得到的结果 Q Inference time frame per second FPS and model size are also very important criterions which should be considered A In Table I we compare actual inference time of regular Conv and CAC models on single TitanX GPU CAC achieves significant inference spee
16、dup while yielding comparable accuracy to regular Conv 展示结果 说明方法有效性 问题案例分析 实验结果类问题 1 可以得到的结果 直接展示 说明方法有效性 2 无法得到的结果 简单说明原因 案例 无法得到的结果 Q It would be interesting to see results toward representation learning semi supervised learning or analogical structure in the latent space A Applications to other tasks such as semi supervised learning are beyond the scope of this paper and will be explored in the future 简单说明原因 问题案例分析 质疑方法重要性或Novelty类问题 1 灵活运用其他审稿人的正面评价予以回复 2 结合文章中的支撑结果解释方法的创新点或重要性 Q I did not f
