Advanced Computing in the Age of AI | Monday, June 24, 2024

Fujitsu Chosen For GENIAC Project To Enhance Reliability Of GenAI in Business Applications 

(issaro prakalung/Shutterstock)

Fujitsu, one of the leading technology and business solutions providers, has been chosen for the research and development project for the enhanced infrastructures for post-5G information and communication systems. This project is part of the Generative AI Accelerator Challenge (GENIAC) initiative by Japan’s New Energy and Industrial Technology Development Organization (NEDO).

The goal of the GENIAC project is to enhance Japan’s capabilities to harness the transformative power of GenAI by bringing together the knowledge of stakeholders in Japan and other countries. Fujitsu will be responsible for R&D on GenAI technologies with a focus on combining knowledge graphs with large language models (LLMs) to enable more reliable use of logical reasoning.

Fujitsu has been working on developing expertise in the development and deployment of GenAI technologies in business operations. The tech giant recently announced the release of Faguku-LLM, a large language model trained on the RIKEN supercomputer Fugaku, one of the world’s fastest supercomputers.

A major challenges for LLM developers is to address the issue of AI hallucinations, a phenomenon that causes GeNAI to create plausible but incorrect or unreliable output. Fujitsu’s research highlights the potential of combining knowledge graphs with LLMs to enhance the accuracy and reliability of GenAI in business applications.

In September last year, Fujitsu introduced new AI mechanisms to improve the reliability of conversational AI. The goal of the new technologies was to provide users with a tool to evaluate the reliability of output from conversation AI models. Building on that success, Fijutsu aims to further enhance the reliability of GenAI.

As part of the GENIAC project, Fujitsu will develop two specialized LLMs: one for knowledge graph generation and another for knowledge graph inference. The combination of the LLMs will allow natural language to be converted into knowledge graphs, which will then be used to derive, aggregate, and deliver the most relevant outputs.

The initial goal is to develop a common pre-trained LLM that will serve as a foundation model for both specialized LLMs. Adding a bilingual corpus to the pre-learning data will enable the common LLM to handle both natural language and knowledge graphs.

With a robust foundational model in place, the more specialized LLMs can then be developed simultaneously. The parallel development will speed up the development process and ensure both models are equipped to handle complex tasks involving natural language and structured knowledge.

If Fujitsu is successful in enabling the realization of specialized LLMs for logical reasoning, it can help create clear, comprehensive, and reliable AI outputs. This could be a game-changer for use cases that require high levels of explainability or compliance, such as internal control and accounting audits in finance, and symptom search and diagnostic support in medicine.

According to Fujitsu, the plan is to offer the new technology to the Japanese market by the end of fiscal 2024. There are also plans to release new technologies via Fujitsu Kozuchi, the company’s dedicated AI service designed to accelerate testing and development of advanced AI technologies.