Foreword : The third part of the closed-loop guide to scientific research: reading methods for academic papers~ When we get a paper, we definitely want the paper to inspire our research work in the next stage. The following are some of the literature reading method for reference
How to read literature
Get involved in a certain field for the first time, read the review, understand the research hotspots, master the key words, and grasp the frontier of the field. When reading a paper, you can start with the following questions:
1. 作者做了什么？
2. 该工作有何意义？
3. 采用何种方法？
4. 该方法具体每一步怎么操作，为达到什么目的？
5. 在做的过程中有哪些难点？作者如何解决？
6. 每张图表都说明什么问题？
How to find improvements
For a certain job, you can think about improvement points from the following aspects:
1. 场景方面：对于应用性研究而言，应用场景的改变意味着新任务的诞生。对于新任务的研究，首先要深入理解现有模型，并指出现有模型为何不能完成新任务。此外，需要做哪些改进适应新任务。
2. 数据方面：所做工作是否需要自建数据集的支撑
3. 模态方面：如文本 ➡️ 图像的转换
4. 模型方面：注意力机制等模块的创新
5. trick方面：如 自然语言生成中的 top-k、top-p 等策略就是在解码策略上的优化
In August 2021, I watched a series of lectures titled "Do the most "sexy" research" by Mr. Hua Guowei from the School of Economics and Management of Beijing Jiaotong University in the series of public lectures on "Postgraduate Scientific Research Quality Improvement". Take notes:
1. 列出方向的核心问题；
2. 画时间轴，掌握研究发展脉络；
3. 建立研究的感性认识；
4. 发现反常现象时，不急于接受否定；
5. ……