How to understand cutting edge AI from scratch

What is AI?
 If you don’t know much about AI technology, you’ll probably think of something like something from an anime or science fiction movie. However, the thing still seems to be a long way off. There are three main areas of AI technology that are currently being researched

  1. natural language processing
     Natural language processing technology is a field that studies the words that humans use in their thinking. Examples include Google Translate, Line chatbots, Amazon Echo, Siri, etc. It is a so-called technology that mimics human thought.
  2. image recognition
     Image recognition technology is a technology that allows computers to recognize the world as it is perceived by humans through their vision. For example, if you enter an image, the technology will recognize it as a dog or a cat.
  3. data mining
     While natural language processing and image recognition have already fixed problems, data mining uses whatever artificial intelligence technology is available for real-world problems.

How to learn AI
 There are two main ways to learn. It’s top down and bottom up. We recommend a top-down method to learn efficiently and quickly.

 The top-down method of learning is one that starts with a real game and works through the details. You decide what you want to achieve and learn the knowledge you need to do so. For example, you can use “least squares” or “linear regression analysis” to “predict future sales”. I’ve gone through the sample code for least squares and regression analysis and can analyze the data I have. At that time, if something that you don’t understand comes up, you can look it up one by one.

 Bottom-up learning is a way to learn from the basics, such as linear algebra, statistical probability, basic machine learning algorithms, deep learning, etc., but when learning the basics, a lot of questions come up, and most of the time we lose time just solving them. What’s more, it’s not necessarily related to the problem you want to solve. This means that a lot of wasteful work will occur. If you have a lot of time to spare and want to enjoy your studies, I think the bottom-up study method is good, but I don’t think it’s suitable for work.

 Top-down learning is a better way to learn and apply than efficiency, so this article will take a top-down approach to learning. The fastest way to learn AI is to present a sample code and have people run the code and look up any concepts or techniques that they don’t understand.

The Shortest Roadmap to the Cutting Edge of AI
 There are many ways to do it, but here I’d like to write some tips that I often use. I’ve also written an example of that below.
Assume a situation where AI is used
 If you research with a sense of purpose, you can find the technology you need faster.
 Example: I want to detect spam comments.

  1. Identify the technical field of the scene you want to use
     The AI technologies that web engineers are often involved in can be broadly categorized into one of the following three categories. It’s natural language processing, or image recognition, or data mining.
     For example: spam comments are natural language processing because they deal with text.
  2. survey the papers of international conferences in the field (applied)
     This step identifies the technology to be used and gives you an idea of the current state of the technology.
     You can find out how to survey from this site.
     Surveying papers from well-known international conferences in natural language processing, image recognition, and data mining. The top conferences for natural language processing are ACL and EMNLP, the top conferences for image recognition are CVPR and ICCV, and for data mining are KDD and ICDE, and the papers of international conferences around AI are almost free and open to everyone.
     The details are summarized in the one-line summary of the conference around AI.
  3. Using the surveyed techniques (Application)
     This step is quite difficult. You must be able to read and understand the paper and have the coding skills to implement its content.
     However, well-known papers often have their source code published on Github, or have been reimplemented and published on Github by other researchers; if it’s published on Github, you’re in luck; if it’s published on Github, you’re in luck. Just go ahead and get the source code, download the dataset, and run it.
     There are a lot of services out there to help with machine learning these days, and simple AI tasks can be found on Kaggle for the most part.
     Example: Source code to detect spam