Artificial intelligence is a broad term covering many technologies, all of which enable computers to display a certain degree of intelligence similar to ours.
The popular use of artificial intelligence is similar to Superman robot in many different tasks. They can fight, fly, and have in-depth conversations on almost any subject. There are many robots in the movie, good and bad, such as vision, Wall-E, terminator, ULTRON, etc. Although this is the ultimate goal of AI research, our current technology is far from reaching the level of AI, which we call universal AI.
Instead, the AI we have today is a subset of AI called narrow AI.
Narrow AI can reach or even surpass the existing human level in some tasks
For example, a few years ago, you may have seen on the news that Google’s AI program deepmind alphago was very good at go. It beat the world champion at that time! However, the program can do nothing but play go. “It certainly can’t play games like pubg or fortnite. It doesn’t even tell you what the current time is.
At present, we are basically exposed to narrow AI, and there are two types of narrow AI. Let’s look at it one by one.
Generally speaking, digital artificial intelligence is also called machine learning
Symbolic AI is also known as old-fashioned AI (gofai) because it has existed for decades. The programmer must manually write all the rules that control the symbolic AI system. As a result, it is difficult to build the right solution. However, it is still used in some use cases where humans need to understand why AI programs make specific decisions in a given situation. For example, if AI judges sentence someone to prison, they must state the reasons for their decision.
Ml machine learning
Ml is relatively new but much more powerful than symbolic AI. Google deepmind’s alphago is an ML system.
In ml, AI programs are not programmers who write all the rules manually, but use lots of examples or data to “learn” what we want to do for ourselves.
It’s similar to how humans “learn” new information. When we want to teach a child the appearance of a dog, we won’t tell him / her that if the animal is short, has drooping ears and wagging tail, it is a dog. Instead, we show children some pictures of dogs. As time goes by, children will naturally understand what a dog is. Ml programs follow the same paradigm.
With the emergence of smart phones and sensors, we produce a lot of data every day, so that machine learning methods now have enough data to be trained. Over the years, the cost of electronic chips such as multicore CPUs and GPUs has also been falling. The proliferation of data created and the availability of cheap hardware are important reasons for the current AI revolution.
Today, there are three main types of machine learning.
1. Traditional machine learning
Traditional ml uses algorithms based on statistical methods to execute ml, among which the most famous algorithms are linear regression, support vector machine, decision tree and so on. Most of the mathematics and statistics of these technologies are decades old and well understood. It was not until the last decade that they were widely known as ml or AI.
To learn the mathematics behind these algorithms, a good reference is the book “the elements of sta tistic learning”. The pythonsklearn and xgboost packages can basically include all of the above requirements for traditional ml using python.
2. Deep learning (DL)
DL has revolutionized the fields of computer vision (CV) and natural language processing (NLP).
In the deep neural network, the multilayer artificial neural networks are linked together, and any mathematical function can be approximated according to the general approximation theorem. Each layer of artificial neural network consists of a linear operation and a nonlinear operation..
By providing the algorithm with a large amount of data about the task we want to learn, we can “learn” the parameters of linear operation. Internally, a learning algorithm called “gradient descent” is used to gradually adjust the parameters until the best accuracy is obtained.
Currently, there are two main Python frameworks for developing deep learning applications: tensorflow and python
3. Reinforcement learning (RL)
In my opinion, reinforcement learning is the most complex of the three ml. Google’s deepmind alphago program beat the best go player in the world, which is an example of RL.
In traditional ml and DL, AI systems learn from past data, while in RL, AI systems learn by taking actions and measuring their returns, similar to training our pet dogs to learn new skills. In games like alphago, the reward is to make decisions to maximize points.
How to choose?
Finally, with all kinds of narrow AI technologies, how do you choose technologies to solve your problems?
First, understand the problem from a business perspective. Then, try various techniques until you reach your business goals. The accuracy rate of 80% using the methods that enterprises can use is better than 99.9% using the methods that enterprises cannot use!
Editor in charge ajx