FPT Software Owns a Deep Learning Patent in the US

A group of 5 researchers from FPT Software (a subsidiary of FPT Corporation) recently received a patent for their technology solution, "Group-equivariant convolutional neural networks for 3D point clouds". The solution can be implemented in training image recognition AI by Deep Learning.

Thanks to its applicability and significantly improved efficiency, the team of experts was awarded one of the world's most prestigious patents by the United States Patent and Trademark Office (USPTO). Operating since 1836, USPTO issued the first US patent signed by President George Washington.

This invention helped enhance the accuracy of training AI to recognize objects from a single angle by improving the neural network's computational performance that separates the objects' features. Additionally, the neural networks were smoothly set up and deployed for numerous parties, making highly functional solutions. At the same time, this way of training AI saved significant resources with fewer input data needed.

Patent granted by USPTO to FPT Software's research team

The researchers came from different departments of FPT Software, such as the global digital transformation consulting department, the AI research department, and graduates from the AI Residency program. Setting the common goal of researching solutions to apply 3D point cloud technology into practice, the experts discovered limitations when combining AI with 3D point clouds.

3D point cloud is a collection of points in three-dimensional space representing a miniature image of objects. It is mainly used in 3D object and image reconstruction. According to the research team, with the evolution of modern 3D cameras, 3D point cloud image stores are increasing day by day.

Using 3D point cloud in deep learning to support AI training has solved considerable problems on 2D images. However, until 2020, world scientists still believed that 3D point cloud applications were at the beginning. Especially if the 3D point cloud data was only from one camera angle, the AI's ability to recognize objects from other angles still needed refinement.

To enhance AI training, the team incorporated an advanced deep learning model called a convolutional neural network, which could decompose input images to produce unique features of objects. As an image recognition application, neural networks are widely used in many industries, including search engines, security, and healthcare. The research team also put their efforts into simplifying the setup and operation of neural networks to bring applications closer to life.

After nearly a year of research, FPT Software experts successfully experimented with a new method of combining neural networks with rotational & equivariant characteristics. The core of this innovation was the infrastructure system that contained available commands. When loaded with data, the 3D Point Cloud will deploy a pre-established neural network that outputs a tensor with all the information of the object in three-dimensional space.

The research team members are graduate students of the AI Residency program

Proudly sharing about the milestone, Mr. Nguyen Xuan Phong - AI Director of FPT Software, one of the five researchers - said: "This success affirms the quality and creativity of exclusive solutions made by FPT Software. Moreover, thanks to our continuous investment in research, we have more confidence to introduce our products and affirm our reputation to the world. Besides, the result lays our core foundation to develop smart solutions in the coming time."

According to USPTO, by 2019, the office had granted more than 4,000 AI/machine learning technology patents per over 60,000 filings each year. Of the AI patents granted worldwide, up to 40% were machine learning, among which Deep Learning and neural networks were in the spotlight of big tech firms, according to WIPO's statistics.

As one of Southeast Asia's leading technology enterprises, FPT Software invested heavily in talent development and AI research. Therefore, it was expected that FPT Software would expand both research and application to projects related to image recognition and deep learning using 3D point cloud data. Besides, the AI Residency training program was one of the opportunities for young researchers to learn from AI experts and to mingle with the world-class AI community.