Careful design of Knowledge Representation allows keeping goals in explicit form that helps the system pursue that goal reliably without deviations and examine conformity along the way.

Discrete structures like partially ordered sets enriched with language and semantic data allow the operator to follow the reasoning trajectory and examine the internal logic of inference process.

Unlike prevailing ML models, e.g. LLMs, our architecture separates reasoning and domain-specific ontology. Architecture is able to incorporate any ontology and can be used for solving a variety of problem classes.

In case of lack of knowledge or contradictions, the system can detect the exact cause of the issue and extract required knowledge from the environment or ask the operator for clarification. This allows the system to work in a semi-supervised manner.

The traditional practice of embedding knowledge in weights, encountered in LLMs, leads to several limitations: the difficulty of locating specific knowledge, costly retraining for adjustments, and expenses associated with processing and acquisition of training data. Through the use of a meticulously designed explicit knowledge representation architecture, these drawbacks can be effectively addressed.

One major challenge with implementing continual learning in Machine Learning systems is expensive and unstable retraining. Noeon Research’s architecture handles new knowledge at the representation level, enabling the system to acquire and solidify up-to-date knowledge and remove obsolete and false information.

Non-agency is an essential property for AI safety, which Noeon Research works on separately from other architectural properties. In order to avoid potentially dangerous behaviour and make the system more amenable, we want to prevent the system from developing agency while retaining reliable goal pursuing abilities.