쓸만한 문구들
배경
카이스트 Scientific Writing 수업의 교과서에서 발췌.
논문 쓰는 순서
Related Work → (Expected contributions) → Results → Method → Introduction & Contributions → Conclusion → Abstract → Title
Abstract
-
Moves
- Background -> Aim -> Method -> Results -> Conclusion (order may change)
-
분량: 0.25 page
-
Note
- 일반적으로 conference 논문에서는 conclusion 부분이 잘 안보인다.
- Aim은 background의 문제 제기 후반부나 method의 전반부에 섞여서 나온다.
- Method나 contributions이 다양하거나 복잡하게 구성되고 내용에 자신이 있으면 method 파트가 더 길어지는 경우도 있다 (Gharbi et al.)
- 가끔 method나 aim이 맨 앞에 오는 경우도 있다. (Chaitanya et al.)
-
Background: 다뤄지지 않은 문제가 무엇인지
- Tense: present or present perfect
- Ex: It has been shown that sth. Sth has shown to be adj to/at sth. Sth offers benefits over sth. Path tracing produces sth. Path tracing requires thousands of samples per pixel to sth. Sth remains a major challenge. Sth has been explored to do sth. Most of these methods rely on sth. Sth has been proved to be useful to do sth. Traditional. Sth limits their power of sth. Sth tends to sth. Some recent works have proposed sth. However, sth. Moreover, sth.
-
Aim: the main objectives/research questions (+ the gap between ideal and reality)
- Tense: present
- Ex: However. While. Sth has not generally has been competitive. Sth has only been used in sth due to sth. To address sth. Our primary focus is on sth.
-
Method: one-sentence summary, 그보다 좀 더 구체적인 방법론 설명
- Tense: past, present, present perfect
- Ex: We propose/present/describe a novel sth. Our technique/algorithm/method do sth. Take sth as inputs. By using sth. To address sth. Our key insight is sth. We adopt sth to sth. To achieve our goal.
-
Results: 말 그대로
- Tense: past, present, present perfect
- Ex: Sth yield state-of-the-art performance. Our results significantly/notably improve sth. Overtake sth. We observe sth. Sth offers sth. Competitive. Compared to sth. Image quality. Sth performs more robustly. Our approach retains sth. Sth is more robust to sth. Our method has the desirable property of sth. Enhance.
Deep learning face representation by joint identification-verification
-
Conclusion: evaluate the results and state the implication of those results
- Tense: present
- Ex: We argue a clear path for making our method run at real-time rates in the near future.
Introduction
- Moves
- General-to-specific
- General problems and solutions → Remaining challenges or pitfalls of prior work → 본 논문에서 사용하려는 방법론의 배경 설명 → Aim & Method overview → Results & Contributions
- 분량: 0.5 page ~ 1.0 page
- Note
- [19TOG_SBMC]는 1, 2, 3 과정을 설명하는 데에 거의 한 줄 씩만 할애했음.
- General problems and solutions: Why your topic matter? 학계에서 중심이 되는 방법론으로 논의를 시작. 해당 중심 방법론을 보완하기 위해 역사적으로 다른 학자들이 제시한 방법론 간략 소개.
- Tense:
- Ex: Perceptually/visually plausible/pleasing.
- For the past few years, deep learning models have been used extensively to solve various machine learning tasks. One of the underlying assumptions is that deep, hierarchical models such as convolutional networks create useful representations of data (Bengio (2009); Hinton (2007)), which can then be used to distinguish between available classes. This quality is in contrast with traditional approaches requiring engineered features extracted from data and then used in separate learning schemes. Features extracted by deep networks were also shown to provide useful representation (Zeiler & Fergus (2013a); Sermanet et al. (2013)) which can be, in turn, successfully used for other tasks (Razavian et al. (2014)). Despite their importance, these representations and their corresponding induced metrics are often treated as side effects of the classification task, rather than being explicitly sought.