Skip to content

ConceptKnowledgeRepresentationRelational

DavidFreely edited this page Nov 7, 2025 · 4 revisions

Representing Relational Knowledge

Let's start out with my favorite simple example of a bit of information. Fido is a dog. Fido is a dog is a particularly useful example because it checks all the right boxes. We know your brain can store this kind of fact. People learn and recall such relationships quickly and effortlessly. Yet despite its simplicity, it's surprisingly complex under the hood.

To represent Fido as a dog, we'll start by adding a synapse between the two neurons with a synapse weight of one. Here's why this single synapse doesn't represent Fido as a dog and what we need to do to fix it. Let's consider that both neurons have numerous other connections, but I specifically want to be able to query my network to ask, what is Fido? And have it contain the information that Fido is a dog. If I fire the Fido neuron, a spike will travel down the axon to the a synapse which will cause the dog neuron to fire because the synapse weight is one.

Since we want the dog neuron to fire only if we're querying what Fido is, let's add a neuron and label it is a. We want to fire the dog neuron only if both Fido and is a are firing. We can accomplish this with a fourth neuron, which I'll label a for convenience. I'll assign it a leakage rate and connect it to the Fido neuron via a synapse with a weight such that no amount of firing Fido will cause A to fire. In this case, a synapse weight of 0.6 is balanced with a leakage rate of 0.3 so that if I set the Fido neuron to fire at its maximum rate, neuron A will never fire. Now I'll add a similar synapse from the is a neuron to neuron A and you can see that if the is a neuron fires while the Fido neuron continues firing, neuron A will fire shortly thereafter. Lastly, I'll connect A to the dog neuron with a synapse weight of 1 so that dog will fire directly after A. Now whenever Fido is firing and is a fires, dog will fire. For those of you with a logic background, neuron A is acting as an AND gate.

  • Source: 2025-07-08 AI, Neurons, and Knowledge of Fido

Clone this wiki locally