Andrew Ng (Andrew Ng) some advice on machine learning, career and reading the paper

作者: Mohamed Ali Habib

Compile: ronghuaiyang

REVIEW

Dry with a lot of very useful.

Introduction

Since you are already reading this article, then you probably already know who one of the pioneers in this field is Andrew Ng, and you are likely to have him on how to build interest in machine learning recommendations career.

This blog summarizes the speech at Stanford University CS230 depth learning courses on YouTube: some suggestions and methods of reading research papers in professional development, link: https:? //Www.youtube.com/watch v = 733m6qBH-jI & list = PLoROMvodv4rOABXSygHTsbvUz4G_YQhOb & index = 9 & t = 0s.

I suggest that you look at this class, very rich in contents. However, I think whether you Kan Bukan, you will find this article helpful. So, here I am trying to summarize these recommendations.

Jump to key points section.

Andrew especially made two major recommendations:

1. Read the research paper: he uses a very effective technique, when he tried to master a new topic in depth study, he would read the research papers.

2. To create a career in the field of machine learning recommendations

Read research papers

How do you efficiently and relatively quickly learn by reading research papers? So, if you want to learn from the academic literature, what you should do, whether you want to learn to build a machine learning system / project of interest, or just to stay on top of things, gain more knowledge and become an in-depth study people.

The following is a list:

1. To prepare a paper list: try to create a list of research papers, including any text or learning resources you have.

2. Go over the list: Basically, you should be in a parallel way to read research papers, meaning a deal papers. Specifically, try to quickly browse and understand every article, rather than read them all, maybe you read each article of 10-20%, maybe this is enough to make your article at hand there is a high level of understanding . After that, you may decide to remove some of the papers of them, or just browse one or two papers, read them again.

He also mentioned that if you read:

5-20 papers (in selected areas, such as speech recognition) => this may be enough knowledge, you can achieve a voice recognition system, but may not be at the forefront of research or make you.

50-100 papers => you might be applications (speech recognition) have a good understanding of the field.

How to read the paper?

Do not read from cover. Instead, we need multiple passes through the paper, here are the specifics of how to do:

1. Read the article title, abstract and drawings: by reading the article title, abstract, key network architecture, and perhaps experimental section, you will be able to concept papers have a general understanding. In depth study, there are many research papers will summarize the entire paper into one or two graphics without having to laboriously read the text.

2. Read Introduction + + Conclusion + Figure skip Other: introduction, conclusion and summary is the author trying to carefully sum up their own place of work, in order to clarify the reviewer why their thesis should be accepted for publication. Also, skip the relevant part of the work (if possible), the purpose of this section is to highlight the work done by other people, and to some extent the work of the relevant work. Therefore, it might be useful to read, but if you're not familiar with this topic, it is sometimes difficult to understand.

3. Read the text, but skip the math section.

4. Read the text, but skip the part does not make sense: excellent research means that we publish something that is on the boundary of our knowledge and understanding in. He also explained that, when you read the paper (even the most influential paper), you may also find that some parts of no use, or no meaning. So, if you read the paper, some of which does not make sense (which is not uncommon), you can skim. Unless you want to master it, it would take more time.

When you read a paper, try to answer the following questions:

1. What the author attempts to complete

2. What are the key elements of this approach is that

3. What you can do

4. Do you want any other references

If you can answer these questions, we hopefully can reflect you have a good understanding of the paper.

It turns out that when you read more articles by practice you will become faster. Because, many authors use when writing papers is a common format.

For example, this is the author of a common format for describing network architecture, especially in computer vision:

 

Paper needs to understand how much time?

For people who are new to machine learning, understanding a relatively simple thesis may take an hour, it is not uncommon. However, sometimes you may occasionally find that you need three hours or even longer to really understand the papers.

Sources paper

There are many great online resources. For example, if you are a novice, blog posts listed speech recognition in the most important papers will be very useful.

With the rapid development of deep learning, a lot of people are trying to keep up with its latest developments. So, you should do:

1. Twitter: Surprisingly, Twitter is becoming an important place for researchers discovered new things.

2. ML subreddit：https://www.reddit.com/r/MachineLearning/。

3. Vital machine learning sessions: NIPS / ICML / ICLR.

4. Friends: find a field of interest to the community or a group of friends, share interesting research papers.

Math section deeper understanding of the text

Try to start again from scratch derivation. Although it takes some time, but this is a good practice.

Practice Code

1. Download open source (if you can find it) and run it.

2. Re-implemented from scratch: If you can do that, then this is a strong signal that you have to really understand the algorithm at hand.

Continuous improvement

The most important thing is to keep on learning, getting better means more stable learning, rather than focusing on a period of time to read a lot of papers. Its rote in a short time, it is better to read the two papers a week starting next year.

Some suggestions for machine learning career

Whether your goal is to find a job (large companies, startups and staff positions), or for more advanced graduate study (and perhaps participate in a doctoral program).

As long as the focus on the opportunity to do important work, your work is seen as a strategy, a useful work to do.

What recruiters want is?

1. Machine learning.

2. Meaningful work: show you qualified for this project work.

For successful machine learning engineer (excellent job seekers) is a very common pattern is developing a T-type knowledge base. Meaning that there is a broad understanding of artificial intelligence in many different subjects, and has a very deep understanding in at least one area.

 

Lateral capacity building

A very effective way to build basic skills in these areas through courses and reading research papers.

Construction of the longitudinal capacity

You can do related projects, open source contributions to research and practice to build it.

Select a job

If you want to continue to learn new things, here are some of the factors that influence your success:

1. Whether you are a great person and / project work together: surrounded by hard-working people it will affect you.

2. In addition to the manager, you will have to focus on working with the team (10-30 people, you will interact most with them), and to evaluate them

3. Do not focus on "brand": the company's brand and your personal experience and not much correlation.

So, if you got a job, ask your team to work together and which do not accept the "join us, then we will form a team," the job offer, because you could do it with a team and you're not interested thing, which is not conducive to their effective evolution.

On the other hand, if you can find a good team (even in an unknown company) and add them, you can actually learn a lot.

Some general recommendations

1. Learn the most: let you learn the most likely to choose something work.

2. Do the important job: to engage in worthwhile projects to promote world forward.

3. Try machine learning to the traditional industries: We have changed a lot in the technology industry, but I think one of the most exciting work may be in the (outside the technology industry) traditional industries, because you can create more there value.

Points

I tried to Andrew's recommendations are summarized as follows:

1. Develop the habit of reading research papers: read the two papers a week as a start.

2. Efficient reading: compile a list of papers, a reading papers, each paper must go through multiple readings.

3. While reading the paper: first read the title / summary / charts (especially) / introduction / conclusion.

4. When trying to understand algorithms: mathematical derivation and try again to practice programming by re-implementing.

5. Try to grasp the latest information, ML meetings by viewing data and other online resources.

6. Construction of a t the type AI knowledge base.
7. Trying to join a good team (in large companies or start-ups), which will help you grow efficiently.
8. Engaged in useful items can help you learn more, pushing the world forward.
9. Try machine learning applied to other industries: medical, astronomy and climate change.

English original: https: //medium.com/@mohamedalihabib7/advice-on-building-a-machine-learning-career-and-reading-research-papers-by-prof-andrew-ng-f90ac99a0182