The full text has a total of 1956 words, and the estimated length of study is 6 minutes
Source: Pexels
Disclaimer: The following is based on the author's observation of the machine learning team, not an academic survey of the industry.
As the core of artificial intelligence, machine learning is a multi-disciplinary interdisciplinary subject, involving multiple subjects such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and so on.
If the reader is a developer, then you may be at least a bit interested in machine learning.
"Well ... the algorithm that can learn predictions by itself, the concept is a bit ... cool."
However, if you really follow the rules to learn machine learning, it is very likely that after two weeks of learning linear algebra and multivariable calculus, the idea that you want to give up will spring up.
This is because most of the introductory materials for machine learning are not for developers, but for machine learning researchers.
Therefore, developers who only want to build products using machine learning will have a headache.
Want to build a product, or do research?
Before the end of the 21st century, machine learning was more or less a research question. The reason is simple, not many companies use machine learning in a valuable way in production.
Therefore, many introductory materials for machine learning are discussed from this research perspective. It first explains the neural network from a mathematical point of view, explaining all the theories behind ML, such as back propagation and opposing networks.
To a certain extent, non-university affiliated materials even follow this model. For example, taken directly from TensorFlow's "Quick Start for Beginners":
Source: TensorFlow
If you are already familiar with the mathematical theory behind machine learning, or are interested in learning it, this method is feasible. However, if you want to mainly use machine learning to build things, it may not be suitable.
Using programming as an inference, this method is like learning to write code by learning assembly first. Few developers learn this way.
Note: If the reader does learn the code by learning assembly, then the author accepts angry comments and is impressed that the reader installs the browser on an obscure Linux distribution.
After learning to code, readers are likely to write hello world by using high-level language. Then, with the introduction of the project a little more complicated, you can learn more and more low-level programming as needed.
This learning mode is very effective among software builders because it can optimize the construction. If readers are more interested in building recommendation engines than publishing papers, the following is how to learn ML.
How to learn ML by building software
Source: Pexels
If the reader is the kind of person who is most likely to achieve results through the course study, then the best resource that can be accessed may be the most active advocate of this top-down, learning-by-doing model, which is the orientation of fastai (Depth of programmer practice) course.
This article is not intended to be a big ad for fastai, but it can be said that several courses organized by its founders, Jeremy Howard and Rachel Thomas, have in-depth research on basic theories related to current construction projects by letting learners actually build Then teach developers deep learning.
Or, if the reader is the kind of person who learns best by doing things by himself, then starting machine learning is the same as other programming fields. Pick a project that sounds very interesting to join. The author here lists a list of beginner-friendly NLP projects.
Assuming the reader learns this way, a good way to become familiar with ML construction is:
1. Determine the goal, such as creating a text auto-completer or license plate recognizer.
2. Find a pre-trained model suitable for the project, GPT-2 or YOLOv3 will be applicable to the aforementioned projects respectively.
3. If you like, you can even use a library like gpt-2 to fine-tune (customize based on the data) the model.
4. Finally, deploy the model as a microservice.
Once the model is deployed as an API, it can be queried like any other web service, and the application built around it.
You can actually see the actual application of this process in this project, which was created by an engineer involved in ML.
In this way, you can learn about different popular model architectures and machine learning methods, and which applications they are suitable for. Understanding ML infrastructure is equally important, and it is essential to build actual products from models.
Once you can use machine learning to build, you must learn this theory
If the reader is an engineer, it may already have some form of authentication, which means that the reader has hashed the password (I hope so).
Did you write a custom hash algorithm when setting the password hash? Did you spend a few weeks learning cryptography? Or did you just use bcrypt?
Similarly, when building the first web application, did it take a few weeks to learn the database? Did you write it from scratch? Or did you use the ORM that came with your favorite framework?
Source: Pexels
The same logic applies to machine learning.
If the reader is the kind of person who likes to build software, start by using machine learning to build software, let the tools, pre-trained models and ML framework abstract the underlying ML theory.
If the reader is still curious or the project needs more complexity, then dig deeper into the machine hood and analyze how it works.
Everyone can always design their own machine learning path.
Leave a comment
Let's share the dry goods of AI learning and development together.
Welcome to the full platform AI vertical self-media "reading core technique"
(Add editor WeChat: dxsxbb, join the reader circle and discuss the newest artificial intelligence technology together ~)
