Machine learning is the cornerstone of modern artificial intelligence, which subverts the traditional programming mode. Machine learning helps to create software that corrects and improves its performance without the need for humans to explain to it how to perform tasks. That's the technology that we use directly today, and many of the innovations that are going to happen, including the amazing suggestions you'll see from websites, digital assistants, driverless cars, analytics software, etc.
What is
machine learning?
Machine learning is software that learns from examples. Instead of writing machine learning algorithms, you train them by providing a lot of relevant data. For example, instead of trying to explain to machine algorithms what a cat looks like, you need to train it by providing millions of pictures of the cat. The algorithm finds repeated patterns in these images and determines for itself how to define the appearance of the cat. After that, when you show a new photo to the app, it can tell if it contains cat elements.
Many people equate machine learning with artificial intelligence. AI, however, is a loose concept that can be applied to anything from complex, rule-based software to uninhabited, human intelligence. In fact, machine learning is a special branch of AI that creates programs based on data rather than rules.
What is supervised, unsupervised and reinforcement learning?
Machine learning has several different styles of algorithms. One of the most popular is supervised learning, which means that you use training algorithms with labeled data to map a set of input objects (usually vectors) to a set of desired output values (also known as supervised signals). The cat example above is an example of supervised learning. Another example is speech recognition, where you can provide a sound waveform to correspond to a written font.
The more samples of algorithms you provide to supervised learning, the more accurate its ability to analyze new data. And that's the main challenge of supervised learning. Creating big data with labeled samples is time-consuming and labor intensive. Some platforms, such as Amazon's Mechanical Turk, provide data tagging services.
In unsupervised learning, another branch of machine learning, has no reference data and no label. In other words, you provide input, but not output. The algorithm sorts out unlabeled data, extracts inferences and finds patterns. Unsupervised learning is particularly useful in situations where humans cannot define hidden patterns.
For example, you allow machine learning algorithms to monitor your network activity. It will then benchmark normal network activity based on the pattern it finds. According to this standard, it will detect and record abnormal network activity.
Compared with supervised learning, unsupervised learning is closer to the process of machine self-learning. However, the problem with unsupervised learning is that its results are often unpredictable. That's why it often requires a combination of human intuition to guide it in the right direction, because it's all self-taught. For example, in the above example of network security, there are many reasons for network activities to deviate from the specification, but this is not malicious. But the algorithm of machine learning is unknown. In the beginning, the human analyst must correct its decisions until it learns to judge anomalies and make better decisions.
Another less well-known area of machine learning is reinforcement learning. In reinforcement learning, programmers define States, expected goals, permitted actions, and constraints. The algorithm tries to combine various allowed actions to understand how to achieve the goal. This is especially effective when you know what the goal is, but you can't define a path to it.
Reinforcement learning is used in many settings. In a more famous case, alphago of Google deepmind has mastered the machine learning program of the complicated board game go. The company is using the same approach to improve the efficiency of the UK grid. Uber is also using the same technology to teach AI agents to play Grand Theft Auto (or, more accurately, let them learn by themselves).
What is deep learning?
Deep learning is a branch of machine learning. Deep learning uses neural networks, a replica of the structure and function of the human brain.
Deep learning solves a major problem in the previous generation of learning algorithms. Previously, with the growth of data, the efficiency and performance of the algorithm platform tend to stagnate. Now, the performance of deep learning algorithm is improving while getting more data. Deep learning algorithm does not directly map input to output, but relies on several layers of processing units. Each layer passes its output to the next layer, processes it, and then passes it on to the next layer. In some models, calculations may flow back and forth multiple times between processing layers. It has been proved that deep learning is very effective in various tasks, including image subtitle, speech recognition and language translation.
What are the challenges of machine learning?
Although machine learning is crucial to the development of future applications, it is not without its own challenges.
On the one hand, the development and deployment of machine learning algorithms largely rely on a large number of computing and storage resources to perform their tasks. This dependency makes them limited to cloud services and big data when they are executed. Therefore, they are more challenging when implementing edge computer integration solutions.
Another problem with machine learning - especially deep learning - is its opacity. As algorithms become more and more complex, it becomes more and more difficult for humans to explain what they are based on and make decisions. In many cases, this may not be a problem. But when you want to make critical decisions about algorithms, it's important to make them transparent and clear.
There are also some biases. Machine learning tends to absorb some habits and tendencies embedded in their training data. In some cases, it is easy to find and eliminate prejudice, while in others, it is so embedded that it is often difficult for human beings to detect it.
However, none of these challenges can prevent AI and machine learning from becoming the general technology of our time (the term has been used for inventions such as steam engines and electricity). No matter where we go, machine learning will have a profound impact on us.
