FOR IMMEDIATE RELEASE No. 3247

Mitsubishi Electric Develops Compact GAN

Offers rapid image synthesis with low computational complexity and reduced memory footprint

PDF Version

TOKYO, January 31, 2019 - Mitsubishi Electric Corporation (TOKYO: 6503) announced today that it has developed a compact GAN (Generative Adversarial Network) based on Mitsubishi Electric's proprietary Maisart®* artificial intelligence (AI) technology. GANs derive from a new machine learning technology that synthesizes photo-realistic images by making two AIs ─ a generator and a discriminator-compete with each other. The computational complexity and memory footprint of the compact GAN is about one-tenth that of a conventional GAN,** a property which enables effective synthesis of the enormous number of images used for the training of other AIs.

* Mitsubishi Electric's AI creates the State-of-the- ART in Technology
** Based on an in-house comparison with our own implementation of a conventional GAN

Overview of GAN and the developed algorithm

Key Features

1)
Reduces the computational complexity and memory footprint of the generator by 90 percent
With a GAN, the AI that synthesizes images is called a generator, and is often realized using a deep neural network requiring significant computational resources and memory. Mitsubishi Electric has developed a novel algorithm that evaluates the significance of each layer in deep neural networks. By removing layers evaluated to be insignificant, the computational cost and memory footprint of the generator can be reduced to about one-tenth of their conventional size** without sacrificing the quality of the synthesized images.
2)
Reduces cost of preparing training images for AIs
Training AI to recognize images requires access to millions or tens of millions of images with diverse variations - one of the biggest challenges of current AI applications, since such data preparation is hugely costly in terms of the time and human resources required. The new compact GAN can synthesize images automatically and rapidly using low-cost devices such as laptops, potentially leading to a significant reduction in the cost of preparing training images for AIs.