An example of the Talking Heads AI developed by Samsung researchers. The AI uses a technique called few-shot learning to create multiple virtual photographs (pictures on the right) with only one original picture (left row). Capture from archive papers.

Artificial intelligence is, of course, an important technology. But it’s not an easy technique to use, you need big data, but it’s not easy to accumulate a lot of data.

Nowadays, good news is being heard about these big data issues. There are a number of different technologies that will provide clues to the solution. Ben Dixon, a software developer and renowned blogger, summed up the technique in his article on Venturebeat.

Hybrid AI
Researchers from the Massachusetts Institute of Technology (MIT) and IBM (IBM) presented a new artificial intelligence called the Neuro-Symbolic Concept Learner (NSCL) at the International Association for Expression Learning (ICLR) held in May. . Looking back at the history of AI development over the last few decades, the approach is largely divided into two. One is the “professional system” pie, which requires us to code the principles of intelligence in clear code, and the other is the “machine learning” wave that lets computers learn the current craze. However, the neuro-symbolic concept learner is characterized by ‘hybrid’ artificial intelligence combining the two.
This AI first extracts the characteristics of the object through learning from some data. He then stores it and combines it with traditional expert system techniques to solve problems. Combining the two, the researchers say, can solve similar problems with a much smaller number of data.

Few-shot learning

Efforts have been made to achieve superior capabilities with little data, one of which is transfer learning . To put it simply, if someone builds a model with big data and publishes it, take that model and train it with some additional data as needed to adapt it to their needs. In the case of the robot employee mentioned in the introduction, the “public employee AI version 1” that has already been made is taken and used for further learning with our store dish. However, this also required hundreds of data and many trial and error processes.
Recently, a technology that can utilize this function with only a few data has attracted attention. This technique is called few-shot learning. Samsung’s researchers introduced a facial animation artificial intelligence called Talking Heads last May. ‘Speaking Head’ captured big parts and features of them through big data learning about facial video. After that, you can create a picture of the person’s different facial expressions and angles with just a few pictures of the person you are seeing. If this feature is introduced to our robotic employees, we can start working with just a few plates. Of course, such a technique would risk further deepening the “deep fake” problem of making video and photographs, as if someone had said and did something.

Let’s make data with artificial intelligence

If artificial intelligence called all-round can not solve big data problem? Of course, it is possible! This is thanks to a new star in artificial intelligence called “genetic hostile neural networks” (GANs). The generative hostile neural network is a technique that creates two artificial intelligences to deceive and artificial intelligence to deceive and fight each other and get amazing results in between. Last month, researchers at the University of Lübeck, Germany, used the technique to “ create ” high-resolution computed tomography (CT) and magnetic resonance imaging (MRI) photographs. [Link] Information about individual medical data is not easy to obtain. The pictures they produced are human data that don’t exist, but they were so realistic that they were enough for other AIs to use for learning. With this, artificial intelligence that needs sensitive data can learn enough by adding a large number of ‘virtual’ data with only a small amount of data.

Ben Dixon didn’t mention it, but what’s indispensable here is federated learning [developed by Google]. This technology does not get the original data from the server, but sends a copy of the AI ​​to the server for study, bringing the learned AI to use the additional learning. Due to the complex AI learning method, it is difficult for humans to know what the original data is based only on the AI ​​they have studied abroad. This technology, if used well, can open the way for the development of superior artificial intelligence without the need to accumulate big data.