November 10, 2017, Tokyo – Preferred Networks, Inc. (PFN, Headquarters: Chiyoda-ku, Tokyo, President and CEO: Toru Nishikawa) has achieved the world’s fastest training time in deep learning by using its large-scale parallel computer MN-1 ¹.

With the size of training data and the number of parameters expanding for the sake of better accuracy of deep learning models, computation time is also increasing. It is not unusual to take several weeks to train a model. Getting multiple GPUs to link with one another for faster training is very important to reduce the time spent on trial and error and verification of new ideas, and produce results quickly.

On the other hand, it is generally known in parallel/distributed learning that the accuracy and learning rate of a model decrease gradually with increased GPUs, due to larger batch sizes and GPU communication overhead.

This time, we have improved learning algorithms and parallel performance to address these issues, and used one of Japan’s most powerful parallel computers with 1,024 of NVIDIA^(R)’s Tesla^(R) multi-node P100 GPUs and leverages Chainer’s distributed learning package ChainerMN ² for training.

As a result, we completed training ResNet-50 ³ for image classification on the ImageNet ⁴ dataset in 15 minutes, which is a significant improvement from the previously best known result ⁵.

The research paper on this achievement is available in the following URL under the title “Extremely Large Minibatch SGD: Training ResNet-50 on ImageNet in 15 Minutes”.
(imagenet_in_15min.pdf)

Based on this research result, PFN will further accelerate its research and development activities in the fields of transportation systems, manufacturing, and bio/healthcare, which require large-scale deep learning.

1 One of the most powerful private supercomputer in Japan, contains 1,024 of NVIDIA^（R）’s Tesla^（R） multi-node P100 GPUs.https://www.preferred.jp/en/news/pr20170920

2 A package adding distributed learning functionality with multiple GPUs to the open source deep learning framework Chainer

3 A network frequently used in the field of image recognition

4 A dataset widely used for image classification

5 Training completed in 31 minutes using Intel^（R） Xeon^（R） Platinum 8160 x 1,600（Y. You et al. ImageNet Training in Minutes. CoRR,abs/1709.05011, 2017）

■ About the Chainer Open Source Deep Learning Framework

Chainer is a Python-based deep learning framework being developed mainly by PFN, which has unique features and powerful performance that allow for designing complex neural networks easily and intuitively, thanks to its “Define-by-Run” approach. Since it was open-sourced in June 2015, as one of the most popular frameworks, Chainer has attracted not only the academic community but also many industrial users who need a flexible framework to harness the power of deep learning in their research and real-world applications.

Chainer incorporates the results of the latest deep learning research. With additional packages such as ChainerMN (distributed learning), ChainerRL (reinforcement learning), ChainerCV (computer vision) and through the support of Chainer development partner companies, PFN aims to promote the most advanced research and development activities of researchers and practitioners in each field. （http://chainer.org/）

■ About Preferred Networks, Inc.

Founded in March 2014 with the aim of promoting business utilization of deep learning technology focused on IoT, PFN advocates Edge Heavy Computing as a way to handle the enormous amounts of data generated by devices in a distributed and collaborative manner at the edge of the network, driving innovation in three priority business areas: transportation, manufacturing and bio/healthcare. PFN promotes advanced initiatives by collaborating with world leading organizations, such as Toyota Motor Corporation, Fanuc Corporation, and the National Cancer Center. (https://www.preferred.jp/en/)

＊Chainer^（R） is the trademark or the registered trademark of Preferred Networks, Inc. in Japan and other countries.