#Work_IN
Hello again! This is another article about the intermediate results of my work on AION, a neuro-symbolic system.
Recently I encountered several architectural problems. The system showed alarmingly poor metrics when selecting the correct end-to-end graph, which forced me to rethink several parts of the project.
I realized that the Id Cluster passed by LS was difficult for the neural systems to interpret. Librarian and the CS AI both struggled with graphs. The graph construction may also be part of the problem, but together these issues produced an awful result: roughly 9% accuracy for the final selected graph. I decided to split Librarian into smaller mechanisms so that errors could be traced to a specific module:
- Field Encoder turns the LS state into a local semantic field.
- Boundary Detector finds the boundaries of meaningful objects and events.
- Concept Composer combines vectors inside those boundaries into one concept vector.
- Concept Resolver compares a new concept with the existing space.
- Bridge Candidate Generator proposes a limited number of possible bridges.
- Bridge Context Reader creates a vector representing the basis of a connection.
- Bridge Form Encoder encodes the type and direction of a connection.
- Bridge Verifier compares competing bridges.
- Claim Assembler combines verified bridges into a separate claim.
- Memory Router and Versioner selects a memory block and stores source, time, and confidence.
It is a lot, but for now this is the only way to trace the error that creates a terrible graph.
I also updated the CS AI mechanism. The slime mold could not handle its task, so I built a hybrid system with graph windows. They find the most similar answer in the concepts and record the surrounding concepts and connections in separate parameters. The window is limited to 20 concepts. Graph-selection accuracy increased from roughly 7.5% to 8.5%. That is very little, but it is still progress.
The hardest part of the system is graph processing: graphs are either built poorly or selected poorly. Maybe I will find a solution, or at least raise answer accuracy to 30% — that would already be a good improvement.
I am also developing Social, Cognitive Workspace, and Experience Space. Social helps the system answer more naturally and maintain small talk. Cognitive Workspace should let it think and solve tasks without an exact match. Experience Space will allow AION to remember mistakes and avoid repeating them, like a person learning through trial and error.
In a few days I plan to release AION alpha in closed access. Because of the graph problems, some mechanisms will be unavailable for now, and AION will answer through simplified versions of LS and CS. Yes, it is a workaround, but I am doing everything I can to keep my promise.
That is all for now. Thank you for reading, and keep following AION — it gets more interesting from here :)