The Last Mile Of Artificial Intelligence

Zheng Lei / Wen

"Can machines think?" The problem of Alan Turing, the father of artificial intelligence, has been partially solved, and machine learning has been able to behave like human beings in some fields, such as playing chess.

At present, machine learning has become an important force to promote the development of industry and society. It can realize the automatic decision-making from e-commerce and advertising to education and medical treatment. Face recognition in the field of computer-based image analysis is a good example. If we have a large number of medical images in hand, we can use them to train the machine to learn to look at new photos and speculate whether there is a disease. Machine learning can also be used in company security systems to determine whether visitors are company employees. However, machine learning has also been criticized. The main reason is that this learning method must be based on a large amount of data. It can even be said that this kind of artificial intelligence must be based on big data. In reality, there are only a few samples available for us to solve. This is an essential difference between intelligent machines and human beings. Human learning is not only based on existing information and knowledge, but also draws inferences from one instance and applies one model to another after appropriate modification. This is transfer learning, which is the next step of traditional machine learning.

In the past ten years, transfer learning has received more and more attention and research in algorithm, theoretical research and practical scene application. This book, written by senior experts in this field, is systematic and comprehensive, including transfer learning theory, automatic transfer learning, small sample learning, lifelong machine learning, etc., as well as achievements in computer vision, natural language processing, recommendation system, bioinformatics, behavior recognition, etc.

Babies first learn how to distinguish their parents, and then use this ability to learn how to distinguish other people. Children can learn from just a few examples and quickly sum up the rules. This ability to learn from small data enables us to use and adjust previous experiences to help solve new problems. In this kind of learning adaptability, human beings are far ahead of intelligent machines. We often encounter a small amount of data isolated from each other and fragmented. Sometimes, due to many restrictions, we cannot collect a large amount of data, such as privacy protection. At this time, machine learning has encountered a bottleneck problem that is difficult to overcome. Transfer learning is a solution to this challenge. This learning mechanism can make the artificial intelligence system more reliable and stable, and it can adopt more complex models to cope with the changes that will appear.

Through transfer learning, knowledge can be reused, so that the acquired experience can be repeatedly applied to the real world. If AI can effectively use transfer learning, we can obtain intelligent machines that will learn for life. This is similar to the trajectory of human evolution, and artificial intelligence scientists have been working in this direction. The ability to transfer knowledge has been regarded as the main cornerstone of AI from the beginning. Analogical learning, case-based reasoning, knowledge reuse and reconstruction, and lifelong machine learning all belong to this category. In the field of education and learning psychology, learning transfer has always been an important topic in the study of effective learning. People firmly believe that the best teaching can make students learn how to learn and adapt the knowledge they have learned to the future.

Let's take an easy to understand example of transfer learning. There are two kinds of road systems in the world: left and right driving. For example, the driver's position in the United States and Mainland China is on the left side of the car, and the car should drive on the right side. In the UK and Hong Kong, the driving position is on the right side of the car and the car is driven on the left side. I live in Shenzhen and am used to driving on the right, but when I get to Hong Kong, I dare not drive. It's hard to change my driving habits. But in the future, we can take the auto driving car, and the transfer learning can be used here. We can find out the common features of the two driving modes, so that the automatic driving system can switch freely. We can see that no matter which side the driver is sitting on, it is always the closest to the center line of the road. This fact enables drivers to "migrate" their driving habits from one direction to another. The key element of transfer learning is to find such invariance between different fields and tasks. Of course, the actual transfer learning is much more complex than this task.

In transfer learning, algorithms are still the core technology, including algorithms based on samples, features, models and relationships. Each migration algorithm corresponds to different knowledge transfer carriers. Text mining is a good application scenario of transfer learning algorithm, which can find useful structural knowledge from text and apply it to other fields. For example, emotional classification, online forums, blogs, social networks and so on have a large number of user generated content, from which it is very important to summarize consumers' views on products and services. For different types of products, different types of online websites, and different industries, users may use different words to express their views with the same emotion. In this situation, transfer learning can be used to train a machine with the ability of human emotion classification. And when AI has finished this last kilometer, many people may be aware of the serious threat it poses.