Traditional AI systems struggle to adapt to dynamic real-time data without constant, costly retraining.
We created AGNs to solve this through a novel pre-trained architecture that learns abstractions from time-varying data, autonomously manipulates these abstractions to derive new concepts for out-of-distribution generalisation, and uses these learnings to establish real-time reasoning without retraining.
How AGNs workAGNs operate on a core principle: changes in time-varying systems start at high frequencies before propagating to lower frequencies. This shapes how the network forms and manipulates abstractions.
The network learns mathematical abstractions from synthetic time series data using category theory for structural understanding, moving beyond simple statistical patterns.
The network leverages proprietary learning algorithms that actively manipulate learned abstractions, exploring potential variations and relationships through symbolic and numeric manipulations to test 'what if' scenarios. This enables the creation of new, related concepts which are key for out-of-distribution generalisation.
The network finally constructs a transparent reasoning layer using symbolic AI, transforming abstractions and learned relationships into cause-and-effect relationships for real-time navigation of complex systems.