Korean researchers develop AI chip to alter mobile device images
Professor Yoo Hoi-jun and his research team with the Korea Advanced Institute of Science and Technology (KAIST) have created a generative adversarial networks processing unit (GANPU) in the form of an AI chip processing GANs with low power and high efficiency, Business Korea reports.
The research was presented at the International Solid-State Circuits Conference (ISSCC) in San Francisco in March.
The AI chip can make fast calculations for mobile device image synthesis and restoration, and uses low power and high efficiency to perform image recognition, inference, learning and determination, according to the researchers.
Generative adversarial networks (GAN) leverage AI and deep learning to generate fake data and tell it apart from real data. The technology is used to create deepfakes. GANs are a versatile technology that can also be used to generate and regenerate images, for image conversion, synthesis and restoration, while conventional AI technology is adopted for object recognition, inference, and voice and facial recognition.
Generative adversarial networks consist of multiple deep neural networks which requires multiple processes to operate high-resolution images. This can be hard on smartphones that have limited capacity and memory.
The new AI chip works on mobile devices and can process both multi and single-column deep neural networks without having to send data to a server. The KAIST team also used it to develop a face correction system with 17 features that can be added or deleted from a photo such as hair, glasses and eyebrows.
The development of a deepfake detector leveraging a GAN by Microsoft in partnership with Peking University was reported earlier this year.