AI inside strengthens AI model development service--Highly accurate learning with few images-ZDNet Japan

　AI inside (Shibuya Ward), which provides AI OCR "DX Suite" that can use artificial intelligence (AI) optical character recognition (OCR) using machine learning on February 28 , held a press briefing to explain their platform strategy. The company reorganized on February 11, and Mr. Tokuchi Watakuchi, who had been the company's representative director, president and chief executive officer (CEO) until then, concurrently serves as chief product officer (CPO).

　In addition to announcing the replacement of the Chief Financial Officer (CFO), the Chief Marketing Officer (CMO) was also newly established. Mr. Takaki Okamura has been appointed as the new CFO, and Mr. Yoshiaki Maedo, who served as the representative director of Apple Japan, has been appointed as the CMO.

　Currently, we are working on preparing a marketplace that will ``realize diverse AI by preparing a place to share AI'' (Mr. Watakuchi). Aiming to realize Furthermore, he pointed out the goal of "shifting from providing SaaS products to AI development, realization of in-house operation, and conversion to a platform that can provide developed services to the market" (Mr. Watakuchi).

The lower limit of images required for machine learning is reduced from 300 to 20

In addition to "DX Suite", AI inside is a cloud-based service that allows you to develop and operate AI models without code. We provide “Learning Center” and “AI inside Cube”, an edge computer that operates AI within the user company. According to Mr. Watakuchi, the Learning Center has made it possible to "confirm (visually) what had to be done using an API with a single click."

　For example, when using AI to check screws with different shapes in the factory, building a learning model and visual observation by humans were necessary, but the preview screen The work efficiency is improved by showing the inference results together with the results. Using the same analogy for the learning data import function, "Uploading pictures of different screws and creating learning data for the Learning Center increases the number of work processes in the factory. Import teacher data created in a different environment. By doing so, the AI ​​learning process can be greatly reduced,” said Mr. Watakuchi.

　In addition, the lower limit of the number of images required for machine learning was previously 300, but this time it has been reduced to 20. "Of course, whether or not 20 sheets are enough depends on the work, but we can achieve a certain degree of accuracy," says Mr. Watakuchi. In addition, the billing system was changed to a pay-as-you-go billing system in units of one second in consideration of cost reduction for users.

DX Suite offers a full-text reading function that is of great interest. "Previously, we were able to read standard and non-standard documents with roughly the same items, but now (with the implementation of this function) we can read anything," Mr. Watakuchi emphasizes. Specifically, by linking the text and the coordinates in the document for the data read by AI OCR, it is possible to accurately read even documents containing mixed handwritten and printed characters.

Focus on collaboration and alliances

ZDNet Japan Articles are summarized by email every morning (free registration)