Our laboratory conducts research and development related to AI orchestration technology, which integrates multiple AI models and data to solve complex problems that cannot be solved by a single AI model.
By leveraging the strengths of multiple models and complementing each other, we expect to be able to build more powerful and versatile AI systems.
We are conducting research on technologies that combine AI models that handle different types of data and tasks as needed to improve and evolve performance, as well as distributed machine learning technologies that safely and efficiently utilize distributed models, data, and computing resources.
We are conducting research on data warehouse/data lake technologies that enable the collection and cross-domain utilization of various data, data mining technologies that discover useful combinations of data, and synthetic data technologies that utilize these technologies to create high-quality training data.
We are developing an AI framework that implements our fundamental technologies and working on specific applications such as smart mobility that utilize this framework. We are also promoting the development of a platform that supports creation and utilization of information assets packaged as AI models, programs, and data sets created through these efforts.
AI models have unique strengths and weaknesses depending on differences in training data and design guidelines. For example, there are various models with distinct characteristics, such as LLM suited for long-form text processing and LLM specialized in specific fields like healthcare, finance, and transportation. Tens of thousands of such models have been developed by companies, universities, and research institutions and are publicly available through model hubs. Amid this landscape, studies to effectively combine multiple LLM to achieve performance and functionality beyond the limitations of individual models, thereby building more powerful and versatile AI systems, are gaining significant attention.
We are reseearching various AI model collaboation technologies such as merging methods that creates a new model by integrating the parameters of multiple models, ensemble methods that combine the outputs of multiple models to derive final results, and cooperation methods that leverage the characteristics of different models to achieve specific tasks. We are also researching distributed machine learning technologies that safely and efficiently utilize distributed models, data, and computational resources, such as federated learning, as well as AI agents that integrate AI models with external databases and services.
Evolutionary multimodal AI for creating highly-accurate multimodal prediction models by combining pre-trained models for each data type (modality).
Federated learning optimized for edge environments, which addresses limited computing resources , device diversity, individual data collection, and unstable networks.
Overview of Event Data Warehouse
We are developing an AI framework that implements our fundamental technologies in MLOps etc. to support safe and efficient data analysis and model development. Based on this framework, we are promoting the development and demonstration of smart services, such as smart mobility, based on predictive analysis of complex real-world situations by integrating various IoT data from different fields.
For more details, please refer to the DCCS of Integrated Testbed at the National Institute of Information and Communications Technology
Examples of application to AI dashcam systems
Cyclical Evolutionary AI Dashcam System