Jamie. ⇄  

“Artificial Intelligence is whatever hasn't been done yet”

A subjective proto-ontology 2021_02_26

The next step is to build a basic, top-level subsumption hierarchy. I want it to be subjective and accurate. By "subjective", I mean "as experienced by a thinking subject". By accurate, I mean that I'm trying to reach pure truth.

Instead of starting from "Everything is an entity" (where "entity" means a part of the world), I'll try to start from "Everything is a perceptual experience", because to a subject, everything, from external stimuli to mental events, exists through perceptual experience.

A simple but surprising consequence is that the past and the future don't exist. There's only a perception of thoughts about the past, and of thoughts about the future.

Usually, the purpose of a top-level ontology is to support semantic interoperability among several domain-specific ontologies by providing a common starting point for the formulation of definitions. The purpose here is different: eventually we want the program to reason about itself.

		
[root] Perceptual experience

	[gen] State

		[gen] TripleStore

		[gen] VectorStore

		[gen] PointStore

		[gen] Native variable (JS)

	[gen] Event

		[gen] State modification

		[gen] Function call (JS)

		[gen] External input

		[gen] Internal input

		[gen] External output

		[gen] Internal output
				
		

• The TripleStore uses LevelGraph for storage and retrieval of triples. Subjects, predicates and objects of a triple are strings. These strings are keys to access vectors from the VectorStore. The TripleStore is where qualitative and symbolic knowledge is stored. It can store structured and unstructured data, answer queries and apply forward chaining rules.
• The VectorStore uses Vector-Object, an N-dimensional vector implementation that's capable of meaning-vector calculations. The VectorStore captures the meaning of things, where meaning is expressed as a multidimensional value. It works a bit like word embedding, except the vectors are transparent (identified vector components) rather than black boxes (anonymous vector components). Vectors are also stored in the PointStore.
• The PointStore uses kD-Tree, a space-partitioning data structure for organizing points in a k-dimensional space. It is able to quickly find the closest vectors to a given meaning, bringing an additional (and important) feature to the VectorStore original toolbox. Its main purpose is to search the nearest known vectors of a calculated vector, which is supposed to provide recognition capabilities.

An important aspect of the global design is the presence of a "feedback loop" that sends descriptions of what's happening inside the program back to its input, so the program can observe its own behavior, predict it, and eventually understand it. This could produce some interesting effects.

LevelGraph Vector Objects kD Tree

The Github repo is open 2021_02_23

Jamie has her own Github repo.

Github