AIの最先端は論文で学ぶ - スマートマニュファクチャリング

This article is recommended for those who

Managers who develop AI businesses, data scientists, machine learning engineers, and anyone studying artificial intelligence

Time needed to finish reading this article

5 minutes

Introduction

nice to meet you. My name is Tsuchiya, and I was assigned to the AI Research Center in December.
Although it is irregular from this month, I will write about AI-related technology here. Thank you.

Now, suddenly, I have a question!

How do you keep up with the latest trends in artificial intelligence?

It is of course important to catch up with various information media such as books, articles on the internet, seminars, news, etc.

The only way to grasp the true latest trends in AI is to learn from papers!

There are many AI researchers in the world, and every day they come up with new methods and try and error how to apply them in the real world. The crystal is the thesis.
Technicians at companies that are developing AI-related businesses need to catch up with them.

Why do we need to catch up with AI trends? As for that, it 's indisputable.
When it comes to AI, efficient methods are constantly being developed by researchers. Whether or not you know them will change the range of possible businesses.

For example, what could not be done without stacking five GPUs this year may be possible next year with a mobile device. Knowing that might even change the business model, and that kind of phenomenon happens a lot in this industry.

Managers who are developing AI-related businesses should take the initiative in having engineers read papers.

Engineers should take the lead in reading papers.

また、弊社マクニカでは、常に論文でのキャッチアップは怠りません。

I will post a summary of the paper on this tech blog in the future, so I hope you will take advantage of it.

Select papers that have been accepted by academic societies

Even if you are told that there are good things about reading papers, you should read the paper first. i don't know how to find I think.
For example, arXiv(https://arxiv.org/), but if you search for "AI" there and find really good information about AI, Almost impossible.

There are many researchers in the world who publish their papers, so these media are updated with hundreds of papers every day.
Therefore, it is as difficult as finding a really good ramen shop in Yokohama to find a truly valuable article in such a medium.

So how do you find articles?

We recommend searching for valuable papers from papers accepted by AI-related academic societies. There are several AI-related academic societies, so I will give an overview.

AI-related academic societies

Although AI-related academic societies have their own characteristics, we have summarized the characteristics of each academic society here.

name	detail	place	season
AI in general
IJCAIMore	The world's top conference not only for machine learning but AI in general	Sweden	July
AAAI	Conference equivalent to IJCAI	Hawaii	Hawaii
JSAI	A Japanese conference called the National Conference of the Japanese Society for Artificial Intelligence	Kagoshima	June
Statistical Machine Learning/Deep Learning
NeurIPS	Top machine learning conference, was called NIPS until last year	Canada	December
ICML	A top conference on par with NeurIPS, which used to focus on experiments, but in recent years has shifted to theory-oriented	Sweden	June
IBIS	Japan's largest machine learning workshop	Sapporo	November
computer vision
CVPR	top conferences in computer vision	America	June
ICCV	Conference on par with CVPR, held every other year	Italy	October

*Researched in December 2018. All conference dates are for 2018. Specialized in natural language processing include ACL and EMNLP.

各々目的にそって、学会を絞りましょう。

例えば、「機械学習やDeep Learningに関する最先端の情報を知りたい」というニーズがあるのであれば「NeurIPS2018」と検索をして、NeurIPS2018に採択された論文を探しましょう。

Or, if you want to catch up on the cutting edge of computer vision, search for "CVPR2018" and read the papers accepted for CVPR2018.

How to use AI-based academic conferences

As for how to use the conference, it is very valuable to actually go, but it is so popular that it is not easy to get tickets.
However, if you look at the society's homepage, you can see the papers accepted by the society and tutorials explaining the AI trends of the year.

Also, some papers have git hub URLs that you can try immediately, so you can easily check how effective they are. In addition, the tutorial may have a video, so it is a good idea to check that as well.

How to read papers effectively

How to read is very important if you read a thesis.

効率的に論文を読むには、論文を読む順番を意識する必要があります。
個人的には、論文は小説と違って大切に読むものではないので、アウトプットを前提に読むと良いと考えています。
私は、現筑波大学学長補佐の落合陽一先生流の論文の読み方を採用しています。

まず、論文を読む順番に関してですが、論文は普通、以下のように構成されています。

1. Abstract → 2. Related work → 3. Experiment → 4. Conclusion

Note that you are not reading from front to back here. Because, although I did an experiment, the result was not well understood as a conclusion. Because there are many things like that.
Therefore, the recommended reading is as follows.

1. Abstract: Get the overview
2. Conclusion: Know what you got first
3. Experiment: Know what led you to that conclusion
4. Related research: Know what other research you need to know to deepen your knowledge

Summary

Here, I explained "how to select papers" and "how to read selected papers" from the point of "learning the cutting edge of AI from papers".

Next, the conference “NeurIPS2018”, which specializes in statistical machine learning and deep learning, was held the other day from December 3rd to December 9th.
In the next article, I will post a summary of NeurIPS2018, so please look forward to it.