Developing deep neural networks' compaction technology, paving the way for applications in self-driving systems and others

“Honestly speaking, I didn't foresee that the day would come when so much attention would be paid to AI technology as it is today. I've already experienced two “ice ages” in AI research, and this third wave surely won't be a transient thing,” Tanizawa says. The biggest reason why he thinks this way is the progress made in “machine learning,” which is the technology that tries to bestow human learning abilities to AI. He says, “Through advanced machine learning, visual information processing has advanced exponentially. In conventional machine learning, when you try to make a computer recognize a cat inside a certain image, you first had to teach the computer what constitutes a cat, or “the conditions for being a cat.” But through advances in machine learning, this task has been automated, and the system of learning itself has been revamped.”

A typical example of advances in machine leaning is “deep learning.” It is a new type of machine learning made possible through “deep neural networks,” which are the multi-stratification of “neural networks.” Inspired by the way information is transmitted by neurons in the human brain, neural networks are mathematical models for mimicking some brain functions on a computer.

Tanizawa is also pursuing research and development with an eye on the future, in which products more familiar to us in our daily lives such as household appliances and other IoT devices are equipped with deep neural networks. This technology is the “compaction technology of deep neural networks that achieves high precision recognition processing on edge devices, which was developed through theoretical analysis conducted in collaboration with the RIKEN Center for Advanced Intelligence Project (AIP).

He explains, “In a conventional deep neural network, there is the problem of edge devices (devices that connect individual communication networks to each other) having difficulty operating because an edge device has limited computing power and memory.” Tanizawa took part in a joint research study at the RIKEN AIP – TOSHIBA Collaboration Center established in 2017, and in 2018, developed a technology that enables a deep neural network to maintain recognition accuracy while reducing the “parameters” arising from the results of a deep neural network's learning by 80%. “This technology enables highly accurate voice and image recognition processing on edge devices. Beyond this, this technology will pave the way for the use of deep neural networks in the image recognition systems of self-driving systems and others,” he says.

Medical technologies, IoT devices, self-driving…
The use of AI will keep on expanding.

AI technology includes deep learning that can find highly accurate, optimal solutions based on huge amounts of data. The use of this technology is expanding – in the medical field (genomic medicine, diagnostic imaging, pharmaceutical development, etc.), in IoT devices for the home such as smart speakers, and in self-driving vehicles. Tanizawa shares his thoughts on the future of AI. “The first thing will be self-driving. By introducing high-performance deep neural networks in self-driving vehicles, the ‘motorized society of the future’ will come faster than expected. Similarly, automation in things such as drones will advance even further. Also, the number of fields wherein AI demonstrates better performance than humans will steadily increase, and AI will replace humans in performing tasks in those fields.”

Tackle AI research in a way
that capitalizes on Japan's strengths.

What kind of role will be expected of researchers who develop AI at corporations? Tanizawa says, “Firstly, they should be able to propose solutions to issues within the company by using AI. Then, they should grasp global trends and recognize Japan's unique strengths. We can say that AI research is now one of the most exciting fields of research. In the research field of AI compaction alone, tens of papers are being published each month by researchers around the world; and in the largest IT companies, there are thousands of researchers. So, it will be meaningless to carry out corporate research in the same way as they do. The important thing is to compete by capitalizing on Japan's unique strengths such as advanced mathematics. Another way is to conduct research on phenomena that are discovered by using real data (which companies are good at) and apply it to society in a way that can be explained theoretically by collaborating with university professors who are good at mathematical analysis. There are still many more challenges in AI research. For example, someone has yet to explain how deep neural networks actually generated optimal solutions. In response to these challenges, I want to carry out work that contributes to the world such as international standards by working in a way that capitalizes on Japan's strengths.”