Introduction of researcher

Sekitoshi Kanai

Distributed Computing Technology Project

Making the current deep learning with AI faster, easier and safer.

I engage in research in the field of deep learning, belonging to a team that works on platform technologies to support “corevo®,” which is the collection of NTT AI (artificial intelligence) technologies. AI is getting a lot of close attention from various fields since AI makes people's life convenient. For example,  image recognition by AI is expected to be utilized for automated driving of cars and machine translation can translate languages instantaneously. The most attractive technology in AI technologies now is deep learning. Technologies that can make time-consuming deep learning faster and easier are required for convenience. In addition, their safety must be secured.

Seeking the AI so that everyone can easily use it

Since deep learning requires time and high expertise to use it, it is not a user-friendly technology yet. Let us take an example of machine translation which has made remarkable progress by the deep learning. We first input a huge number of English sentences data into a computer. Like human beings, this process is called “learning” and entering Japanese sentences as well as English allows a model to learn the relationship between both languages automatically. For example, instead of letting the computer learn each word like “apple = ringo”, entering “this is an apple” allows the computer to learn bilingual sentences to give back “kore ha ringo desu”. If the model learns a lot of those bilingual sentences, it can learn automatically including meanings of words. And then, the model will be able to highly and accurately translate even sentences that are not included in the data. It takes a long time for the model to only learn and usually a week or two to learn sentence data ranging from several million to hundreds of millions of words. Because there are many parameters to be tuned in a trial-and-error manner for the learning, we need knowledge and skill to tune them. If the difficulties are lessened and everyone can implement deep learning easily and stably, the spread of AI in the world will be accelerated dramatically.

Adversarial attack as a new “risk”

AI is not free from the risk management, to which the world is paying attention. New risk management about utilizing the AI can be required. It is reported that images injected a particular imperceptible noise is misclassified by AI. This noise can be a risk for deep learning applications. For example, If an autonomous car recognizes a sign by deep learning, it is possible to erroneously recognize the sign and to lead to incorrect driving by an above-mentioned noise. This is called an adversarial attack, and technologies to prevent it are still being studied.

Group of members to study diversified fields.

Members in a wide range of research fields gather in the SIC based on software. While proceeding research individually on a daily basis, we regularly make opportunities for group meetings to share our efforts with each other. While a barrier is inseparable amid the researching, we sometimes find clues as a result of the meeting. Logical thinking is required for communication, and integration of what we are doing is required for preparing documents. The preparation process allows us to objectively observe our own research. Furthermore, through discussions with the members, we may find our weak points or unsatisfied parts of our theories. When you are thinking alone and fail to find a solution, the situation where you can discuss easily is an advantage of the SIC.

I have been joining the company for five years. I have been trying to find my research theme based on what I can do but from now on I would like to improve the research that can meet the needs considering what business companies need and what my output can produce.


Sekitoshi Kanai

Distributed Computing Technology Project

Career background

Joined Nippon Telegraph and Telephone Corporation.
Have belonged to the Software Innovation Center.
Had engaged in research and development on deep learning.
April 2018
Proceeded to a doctoral course at Keio University.
Research presentation

Research presentation

  • Sekitoshi Kanai, Yasuhiro Fujiwara, and Sotetsu Iwamura. "Preventing gradient explosions in gated recurrent units." Advances in Neural Information Processing Systems (NIPS) 2017.

  • Sekitoshi Kanai, Yasuhiro Fujiwara, Yuki Yamanaka, and Shuichi Adachi. "Sigsoftmax: Reanalysis of the Softmax Bottleneck." Advances in Neural Information Processing Systems (NIPS) 2018.