Representing Knowledge
Posted: Mon Mar 20, 2006 7:51 pm
I have recently done a lot of research on Knowledge systems, and how knowledge is represented. From this, I created a neural network program, that simulates the way concepts and associations are stored in the brain. Now I am faced with the challenge of how to actually populate such a system, because one of the things that becomes apparent is that there cannot be any "assumptions" made... everything has to be put in context.
The program creates "neurons", which have several attributes:
1) Context. Each neuron has a linkage to a context neuron, like a "parent concept". Spatial location is highly contextual. For example, your pillow is on the context of your bed, which is in the context of the bedroom, in the context of the home, in the context of the street in the context of the neighborhood,... etc, on and on, right up to the planet Earth, in the Sol solar system, and whatever galactic arrangments occur beyond that. It is how one would "zoom in" to the neuron, along a specific, and unique, path from the origin of all.
In the program, each neuron has only ONE context it can be in. A pillow cannot be on the bed and on the sofa at the same time.
I will admit that I still have some more conceptual work to do on the "temporal context" within the program design, because you CAN have the same pillow on the bed and on the sofa, at different TIMES. Right now, I am addressing Time thru "association", described later. But I feel there is a better way; just cannot see it yet.
"Context" is often referred to as "scope" in other knowledge systems.
2) A unique "base name" that allows us, as humans, to attach some sort of conceptual meaning to the neuron. Base names must be UNIQUE within their context. For example, "Mercury" has multiple contexts: a planet, an element, an automobile, etc. The base name of "Mercury" would work in all three examples, because they would be within the context of Astronomy, Physics and Vehicles, respectively.
3) Associative references. In order to address the non-local characteristics of connectivity, neurons can also associate other neurons in a "player--role" model. For example, in the context of "Human relationships", one has a "Family" assocation. Within that association, there are players and roles -- mother, father, son, daughter, etc. The "player" linkage of a neuron refers to what is being acted on, and the "role" links the action.
4) Facets of Information. Neurons can contain specific facets of knowledge. For example, a "Bruce's Birthday" neuron would have a facet of the date and time I was born. I have included facets for the concepts of text, numbers, date-times, URLs and comments. These facets are used for dynamic information, analogous to "short term memory" in the brain. Common usage of a facet on a computer system would be a username, password, email address, etc., which can be altered, without changing the base concepts.
I realize that this is a "short cut", and that all information can be represented via the original concept/association model. I did actually attempt such a program, but the overhead became phenomenonal, since even things like character sets had to be represented, and associations to form words in cryllic systems. This is here more for practicality, than conceptual design.
The neural network program I have created (and will have online shortly), simply creates and stores neurons. But it knows nothing to start... the next challenge is to take the concepts of the Reciprocal System, and give them context, association and facet. And believe me, it points out errors in logical deduction VERY quickly!
I will be following up with my ideas on how to create an RS knowledge structure, based on this model.
The program creates "neurons", which have several attributes:
1) Context. Each neuron has a linkage to a context neuron, like a "parent concept". Spatial location is highly contextual. For example, your pillow is on the context of your bed, which is in the context of the bedroom, in the context of the home, in the context of the street in the context of the neighborhood,... etc, on and on, right up to the planet Earth, in the Sol solar system, and whatever galactic arrangments occur beyond that. It is how one would "zoom in" to the neuron, along a specific, and unique, path from the origin of all.
In the program, each neuron has only ONE context it can be in. A pillow cannot be on the bed and on the sofa at the same time.
I will admit that I still have some more conceptual work to do on the "temporal context" within the program design, because you CAN have the same pillow on the bed and on the sofa, at different TIMES. Right now, I am addressing Time thru "association", described later. But I feel there is a better way; just cannot see it yet.
"Context" is often referred to as "scope" in other knowledge systems.
2) A unique "base name" that allows us, as humans, to attach some sort of conceptual meaning to the neuron. Base names must be UNIQUE within their context. For example, "Mercury" has multiple contexts: a planet, an element, an automobile, etc. The base name of "Mercury" would work in all three examples, because they would be within the context of Astronomy, Physics and Vehicles, respectively.
3) Associative references. In order to address the non-local characteristics of connectivity, neurons can also associate other neurons in a "player--role" model. For example, in the context of "Human relationships", one has a "Family" assocation. Within that association, there are players and roles -- mother, father, son, daughter, etc. The "player" linkage of a neuron refers to what is being acted on, and the "role" links the action.
4) Facets of Information. Neurons can contain specific facets of knowledge. For example, a "Bruce's Birthday" neuron would have a facet of the date and time I was born. I have included facets for the concepts of text, numbers, date-times, URLs and comments. These facets are used for dynamic information, analogous to "short term memory" in the brain. Common usage of a facet on a computer system would be a username, password, email address, etc., which can be altered, without changing the base concepts.
I realize that this is a "short cut", and that all information can be represented via the original concept/association model. I did actually attempt such a program, but the overhead became phenomenonal, since even things like character sets had to be represented, and associations to form words in cryllic systems. This is here more for practicality, than conceptual design.
The neural network program I have created (and will have online shortly), simply creates and stores neurons. But it knows nothing to start... the next challenge is to take the concepts of the Reciprocal System, and give them context, association and facet. And believe me, it points out errors in logical deduction VERY quickly!
I will be following up with my ideas on how to create an RS knowledge structure, based on this model.