25 Years of Expertise Guiding AI to the “Next Dimension.”
Over the past 25 years, Mindware Research Institute has been at the forefront of data mining utilizing Self-Organizing Maps (SOM) and building Bayesian networks. Building upon this proven track record, we have evolved towards a unique architecture designed for the next-generation AI agent era. Here, we introduce our core technologies that unlock the true value of ultra-high-dimensional data and Large Language Models (LLMs) to solve complex challenges.
1. GNG+MST Conceptual Network Model
~ Next-Generation Data Structuring Technology for Ultra-High-Dimensional Data ~
To adapt to the AI agent era, we have shifted our focus from traditional SOM families to GNG (Growing Neural Gas), a technology that allows networks to grow more flexibly. Utilizing GNG, we developed “ConceptMiner,” a proprietary engine that automatically constructs complex conceptual network models, enabling advanced analysis and visualization of ultra-high-dimensional data.
2. Self-Organizing Knowledge Base
~ A Next-Generation QA Platform Seamlessly Integrating Corporate Documents ~
A system infrastructure that transforms vast amounts of dormant corporate documents into a knowledge base, providing flexible and accurate responses to natural language queries.
- LLM Wiki Compilation: Compiles knowledge from individual documents using our proprietary “LLM Wiki” method, enabling highly practical and precise responses.
- Large-Scale Integration and Optimal Routing: Cross-integrates multiple LLM Wikis to build a massive, unified knowledge base. User queries are instantly analyzed by our GNG+MST Model and accurately routed to the most appropriate LLM Wiki.
3. Analogical Reasoning (Connected Concepts Model)
~ “Associative” Knowledge Exploration from Different Perspectives ~
A reasoning model that goes beyond simple keyword searches, enabling “associative knowledge exploration” akin to human thought. By creating multiple text strings for each entity, we concurrently build multifaceted GNG+MST models. In response to a user’s query, nodes are first matched in [Model A] to extract related information. By tracing those entities and exploring [Model B], the system can draw out related information from entirely different angles, uncovering new insights and inspiration.
4. GNG+KG (Self-Organizing Knowledge Graph)
~ A Low-Cost Approach to Efficiently Organizing and Integrating Massive Information ~
“Knowledge Graphs (KG),” which structure relationships between entities, are becoming widely adopted in AI applications. However, exhaustively modeling explicit relationships manually entails massive design costs. We solve this problem using GNG. The system automatically generates a local knowledge graph for each GNG node, integrates them within clusters, and finally achieves global integration. This is a breakthrough technology for efficiently and rationally organizing large volumes of disparate information.
5. GNG Episodic Database (Associative Long-Term Memory)
~ Enabling “Learning from Experience” and “Instant Recall” for AI Agents ~
For autonomous AI agents to operate appropriately, they need the ability to instantly and accurately recall past experiences (lessons learned) in alignment with the current context. Mindware Research Institute solves this challenge using GNG technology. The system automatically learns episodes in tandem with the AI’s “short-term memory” and anchors the results into a database (long-term memory). AI agents can instantly gather and integrate relevant long-term memories according to the situation to execute the optimal actions.