A Considerate Robot Brain: A Neural Network Approach

Masafumi Hagiwara (Professor, Department of Information and Computer Science)

It is said that our lives have become much more convenient and comfortable with smartphones and the internet. But is that really true? Current information devices will give you a response if you input the correct information. However, they do absolutely nothing beyond what they are instructed to do, and their responses are cold, or rather, bland and dry. When you think about what's missing, you realize it's a human touch.

Humanity began with the use of tools and went on to invent engines, automobiles, airplanes, and more. These are, so to speak, extensions of the capabilities of human hands and feet. On the other hand, the invention of semiconductors and computers has made it possible to expand the capabilities of the brain. Needless to say, in terms of computational ability, even a calculator from a 100-yen shop far surpasses human capabilities. Pardon the personal anecdote, but the "portable telephone," which was a complete fantasy when I was a graduate student researching communications, has become a reality. Moreover, thanks to smartphones, we now live in an age where we can carry a machine more powerful than the large computers of the past in our pockets.

However... can these high-performance information devices do the things that humans do casually and without thinking? As long as the password is correct, a smartphone will judge the user to be the owner, no matter how different their voice is. A human would immediately know that it's clearly not the owner. Furthermore, between humans, we can sense another person's fatigue from their facial expression or voice and show consideration for them.

In our laboratory, we aim to build a robot brain that has a human touch and can be considerate. Our model is the brain that we possess. We are taking an approach that simulates the workings of a biological brain—that is, a neural network approach.

A neural network is what is called a "nerve circuit network" in Japanese. It is a network formed by the connection of many nerve cells (neurons), which you may have learned about in high school biology. A simple representation is shown in Figure 3. The circles in the figure represent neurons, and it is said that there are tens of billions of them in the human cerebrum. This may seem like a huge number, but let's compare it with the latest processors. Processors equipped with several billion or more transistors are already on the market today. Of course, neurons and transistors differ in function. However, in terms of being basic elements for information processing, we are in an era where a few dozen of the latest processors can surpass the total number of neurons in the human cerebrum. In other words, the artificial brain, which may feel like a fantasy, has actually entered an era where it is fully achievable from a hardware perspective.

It is said that humans have five senses, but in our laboratory, we are conducting research focusing on the following three types of information processing: visual information processing, which handles input from the eyes; language information processing, which makes full use of words; and Kansei/affective information processing, which deals with emotions and sensibilities, the core of humanity. And we are conducting research aimed at integrating these.

(1) Visual Information Processing: Currently, methods using neural networks that simulate the functions of biological brains are attracting significant attention worldwide. Until now, pattern recognition has used unique methods for each field, such as speech recognition, image recognition, and character recognition. However, the deep learning neural networks proposed in recent years not only have excellent characteristics but also possess the generality to be applied to multiple fields and have great potential for future development.

In our laboratory, for example, we are considering the following development methods. First, let's say the input visual information is recognized. But this is merely a conversion from an image to a string of text. Therefore, we perform associative processing on the recognition result, using language information to link it to various meanings and images. Then, we conduct information processing that connects it with related knowledge, past experiences, and even emotions and sensibilities. This is similar to how, for example, when we see a fresh melon, we might think it looks delicious or beautiful, or consider it expensive, with various things coming to mind depending on the scene and situation.

(2) Language Information Processing: Our laboratory has been conducting research on language processing using neural networks for a long time. In 2013, we received the Best Research Award from the Japanese Neural Network Society (Japan's academic society for neural networks) for the development of a language processing neural network like the one shown in Figure 5. The major features of this neural network are its ability to handle complex input sentences, its automatic consideration of deep cases, its capacity for associative processing via the dictionary network at the lowest layer, and its ability to answer questions using these functions.

(3) Kansei/Affective Information Processing: We are developing applied systems that leverage the analysis of the relationship between visual/language information and sensibilities/emotions. Examples include support for interior layout, support for flower arrangement (kadō), support for creating 3D characters aimed at understanding "cuteness" (kawaii), and estimating the impression of a newspaper article from its headline.

(4) Handling Common Sense: Processes that are obvious and simple for humans are often surprisingly difficult for computers. One reason for this is the handling of common sense. In the research field known as artificial intelligence, it has long been said that the automatic acquisition and handling of human common sense is extremely difficult. Our laboratory, which aims to realize brain-like information processing, is also conducting research aimed at this automatic acquisition of common sense. For example, it is becoming possible to make common-sense judgments such as "helping a person in trouble" -> very good, and "walking while looking at a mobile phone" -> very bad. In this way, we are aiming for a robot brain that can, in the future, correctly judge events it sees, help people, and be considerate.