（Cover and body images are from the internet）
Machine learning may be a “familiar yet strange term”. You often hear people mention it, but its specific meaning is not very clear. When we talk about machine learning, what are we actually discussing? This article will organize the discussion from four aspects:
/ What is Machine Learning / Why Machine Learning is Important / Application Areas / Common Methods /
This structure is inspired by an article from SAS titled “Machine Learning: What it is and why it matters”[1]. The article includes my own thoughts and summaries shared with everyone, and any omissions are welcome to be pointed out.
What is Machine Learning
Machine learning, as a branch of artificial intelligence (AI), is essentially a method of automatically building analytical models based on data. Its core idea is that the system learns useful information from data through mathematical formulas, thereby completing certain predefined tasks, such as pattern recognition, numerical prediction, decision-making, etc. In this process, human involvement is minimized, and the machine is the main entity learning and completing tasks.
Machine learning originated from the curiosity of AI scientists: can computers learn things automatically from data? The criterion for “learning things” is whether the computer can still effectively and stably complete the expected tasks when new data is input. For a long time, machine learning was more of a toy for mathematicians. The algorithms have long existed, but practical application scenarios were limited. However, in recent decades, due to the explosion of data volume (thanks to the proliferation of various sensors and networks) and computing power (thanks to the development of underlying chips), machine learning has entered a period of vigorous development.
Machine learning methods have a natural dependence on the quantity and quality of data. The larger the data volume and the higher the quality, the better the results are typically obtained. Machine learning is not omnipotent; without good data, it is impossible to achieve good results. There is a saying in the industry: “Junk in, junk out”, meaning that if low-quality data is fed into a machine learning system, only poor performance results can be obtained. Many people unfamiliar with machine learning exaggerate its role or hold overly high expectations, often without realizing that to fully utilize a machine learning algorithm’s capabilities, sufficient data must first be collected and carefully selected and processed; otherwise, it’s like building a house on the sand. From a theoretical perspective, machine learning is essentially mathematics, related to subjects such as probability theory, information science, and logic; from an engineering perspective, machine learning is merely a tool that dances with data, being one part of the entire system engineering from data collection to analysis and testing.
Why Machine Learning is Important
I personally believe that the most significant meaning of machine learning lies in its role as a tool to greatly enhance productivity. In the era before big data and machine learning, production and life were guided by expert knowledge. Machine learning replaces the role of experts through the power of data. For example, if an imported production line breaks down, only engineers from the country of origin can be called for repairs, while experienced engineers often require decades of experience accumulation. However, we can now use machine learning methods to collect a large amount of data on normal and abnormal operations of machines, and it is hoped that a computer program can determine the cause of the fault and provide repair methods. More importantly, this computer program can be infinitely replicated and used by manufacturers everywhere. Another example is driving, which is also a form of expert knowledge, but this expert knowledge is gradually being replaced by autonomous driving systems. Machine learning systems trained on vast amounts of human driving data have developed the ability to drive autonomously in specific areas such as highways and towns. In the future, our hands can be freed from the steering wheel, and the time wasted on commuting will be returned to us for more meaningful activities. Moreover, the efficiency of intercity and interstate logistics will be greatly improved, as autonomous driving systems can operate tirelessly on highways or railroads 24/7. Additionally, urban traffic congestion may be alleviated, as machine learning systems can reasonably schedule each vehicle’s driving path and speed. Another example is weather forecasting, which is also a form of expert knowledge, but with the help of large amounts of data and machine learning, it is now possible to predict future weather more accurately from historical data.
Similar examples of machines replacing experts abound, reducing production costs and improving production efficiency to varying degrees. From waterwheels for irrigation, windmills for milling, to the steam age and electrical age, the core driving force behind human social progress has been the continuous improvement of productivity. Today, machine learning technology has become an important tool for enhancing productivity in the information age. It will be continuously standardized to the point where it exists around us like air.
Application Areas
In addition to the industrial production, transportation, and meteorology mentioned earlier, machine learning has been applied in many aspects of society. In the financial sector, there are mainly two directions: one is to identify the most valuable investment opportunities and timing, and the other is to prevent fraud or high-risk transactions; governments often use machine learning to identify and track identities or prevent identity theft; in the medical field, machine learning is used to assist doctors in making diagnoses (this is also a typical example of replacing expert knowledge); in the retail sector, businesses use machine learning systems to learn about users’ consumption habits to design more targeted recommendations, marketing, and supply chain management strategies… and so on.
Common Methods
Machine learning methods are mainly divided into three categories: supervised learning, unsupervised learning, and semi-supervised learning. Supervised learning refers to the use of labeled data where each input corresponds to a correct output. The algorithm iterates itself by comparing its output value with the correct output value, thus achieving a certain level of precision. Unsupervised learning uses unlabeled data where each input does not have a pre-defined “correct” output. This type of algorithm generally extracts useful information by mining the structural features of the data itself. Semi-supervised learning, as the name suggests, uses partially labeled data to train machine learning algorithms, typically in situations where it is difficult to obtain a large amount of labeled data. One branch that has gained significant attention in academia this year is reinforcement learning, which has promising prospects in the field of robotics.
For specific problems, the choice of machine learning algorithms is often different. Here is a flowchart that can help everyone understand the differences between various specific algorithms so that they can choose according to their needs: Which machine learning algorithm should I use?[2]
References and Learning Materials
[1]https://www.sas.com/en_us/insights/analytics/machine-learning.html
[2]https://blogs.sas.com/content/subconsciousmusings/2017/04/12/machine-learning-algorithm-use/
Finally, I would like to share some materials in the field of machine learning for reference.
– Coursera Machine Learning by Andrew Ng
– Pattern Recognition and Machine Learning [3]
– Machine Learning [4]
I personally believe that starting with online courses (the first one listed above) or blog articles to quickly understand technical details, then doing some small projects, supplemented by systematic reading of classic textbooks (the second and third ones listed above) is a relatively efficient learning method.
Compiled by: Yang Haijun