HOME > Research Results > Research Highlights, Development of AI That Does Not Require Collection of Large Amounts of Natural Images

Research Highlights, Development of AI That Does Not Require Collection of Large Amounts of Natural Images

Department of Information Technology and Human Factors

Development of AI That Does Not Require Collection of
Large Amounts of Natural Images

AI automatically learns from mathematical formulas and achieves recognition accuracy equal to or even exceed that of manually labeled image dataset

KATAOKA Hirokatsu, OKAYASU Kazushige, MATSUMOTO Asato,
YAMADA Ryosuke, SATOH YutakaArtificial Intelligence Research Center

Pre-trained AI models without any natural images

The world's first method to build an artificial intelligence (AI) image recognition model (pre-trained model) using a large-scale dataset automatically generated from mathematical formulas, without using any natural images for pre-training, has been developed, solving the problems of commercial use of large numbers of natural images used by AI in training, ensuring their privacy, and reducing the cost of manual labeling. In addition, the new method achieves image recognition accuracy equivalent to or better than the current method that uses natural images and human-judged teacher labels in some tasks. It is expected to be applied in the construction of AI for various environments that require images.

“Large amount of natural images" barrier to AI implementation

Large-scale image datasets have been used to train image recognition AI, but building such datasets takes time and effort, and the resulting datasets have data transparency issues such as mislabeling, privacy issues, and copyright issues, which hinder their commercial use. Thinking that not using natural images is the method that basically overcomes the above-mentioned issues led to the conception of this study. There is a need to develop a pre-training method not using natural images that achieves the same or better recognition accuracy than conventional models.

Fractal-generated image patterns outperform other automatically generated datasets

In this study, image data and teacher labels were generated fully automatically from mathematical formulas, and image datasets were constructed. When the image recognition AI was trained by the dataset generated by fractal geometry, a mathematical formula that is said to be forming parts of natural objects, the recognition accuracy was close to that of conventional learning using natural images and human-supplied teacher labels. Furthermore, when recognizing images, since AI mainly identifies objects by focusing on the contour component of fractal geometry, we also constructed an image dataset with a mathematical formula that generates radial contours. With these two image data-sets as bases, after repeated improvements, the recognition accuracy of general object images (ImageNet) by an image recognition AI trained on each of these two image datasets was higher than that based on natural images (fractal geometry: 82.7%, contour shape: 82.4%, natural image: 81.8%).

Expand range of applications through release of trained models

If the transfer accuracy of pre-training without using natural images improves, it will be possible to replace the ImageNet dataset conventionally used, thereby protecting privacy and reducing the labeling cost. The development of "general-purpose pre-trained models" that can serve as a foundation for any task without the use of natural data or human-determined teacher labels is expected to be useful in building AI in a variety of environments, including the medical field, traffic scene analysis, and logistics sites, where high performance is required. We will construct and disclose pre-trained models assuming all scenes and industrial tasks, and will proceed with the project so that more users will be able to develop AI without concern about rights and ethical issues.

Contact for inquiries related to this theme

Computer Vision Research Team, Artificial Intelligence Research Center

KATAOKA Hirokatsu, Chief Senior Researcher

AIST Tsukuba Central 1, 1-1-1 Umezono, Tsukuba, Ibaraki 305-8560, Japan

E-mail: airc-info-ml*aist.go.jp (Please convert "*" to "@".)

Web: https://www.airc.aist.go.jp/en/cvrt/

▲ ページトップへ