The main suggestions made by ndrew:
Pay attention to paper reading
The specific steps are:
1. Write a list of essays: Try to create a list of essays, including any text or learning resources you have.
- Go through the list: read papers in a parallel way, that is, process multiple papers at the same time. Specifically, try to quickly scan and understand each article instead of reading it all. Maybe you read 10-20% of each article, but this is enough to give you a high level of understanding of the article at hand. After that, you might decide to delete some of these papers, or just browse one or two papers and read them through.
2. Don't read from beginning to end. Instead, you need to traverse the paper multiple times. Here's how to do it:
- Read the article title, abstract and diagram : By reading the article title, abstract, key network architecture diagram, and perhaps the experimental part, you will be able to have a general understanding of the concept of the paper. In deep learning, there are many research papers that summarize the entire paper into one or two graphics without the need to read through the full text.
- Read introduction + conclusion + figure + skip others : introduction, conclusion, and abstract are where authors try to carefully summarize their work in order to clarify to reviewers why their paper should be accepted for publication.
In addition, skip the relevant work part (if possible). The purpose of this part is to highlight the work done by other people, which is related to the author’s work to some extent. Therefore, it may be useful to read it, but if you are not familiar with the subject, it can sometimes be difficult to understand. - Read the full text, but skip the math part.
- Read through the full text, but skip the meaningless parts : Excellent research means that what we publish is on the boundary of our knowledge and understanding.
He also explained that when you read a paper (even the most influential paper), you may find that some parts are useless or meaningless. Therefore, if you read a paper and some of its content is meaningless (this is not uncommon), then you can skim it first. Unless you want to master it, spend more time.
3. When you read a paper, try to answer the following questions:
- What the author is trying to accomplish
- What are the key elements of this method
- What can you do
- What other references do you want
- If you can answer these questions, it means you may have a better understanding of the paper.
It turns out that when you read more papers, you will become faster with practice. Because many authors use a common format when writing their papers.
Have a deeper understanding of the mathematics part of the text and
try to re-derive it from the beginning. Although, it will take some time, but it is a good exercise.
Code exercises
Download the open source code (if you can find it) and run it. Re-implement from scratch: If you can do this, then this is a strong signal that you have truly understood the algorithm at hand.
4. Continuous improvement The
most important thing is to keep learning. Getting better refers to more stable learning, rather than focusing on reading a large number of papers over a period of time.
Instead of memorizing in a short period of time, it is better to read two papers a week starting next year.