Memory – Semantic Networks

Why study memory?

We want to know how to access knowledge in order to solve problems, and to do that we need to understand how knowledge is organized in memory.

Two Examples

Example 1

“Jack got a rope.

He wanted to commit suicide.”

(From Charniak and McDemott, Artificial Intelligence, chapter 10.)

Why did Jack get a rope? Because he wanted to hang himself. Answering this question requires accessing one particular bit of knowledge out of the huge knowledge base we have.

Example 2

Does AB look more like C1 or C2?

Answer: C2

(From Winston, Artificial Intelligence, 3^rd Edition, chapter 2.)

These two examples are very different but use the same memory representation: semantic networks.

Semantic Network Representations

For A and B in Example 2:

Four parts of the semantic network representation:

Lexical – vocabulary	nodes, links
Structural	node à link à node
Semantic	node means object link means relation (in this case)
Procedural	Read: Given node, what is link out? Given node, what is link in? Given link, what is preceding node? Etc. Write: Create new node linked in a given way.

Part of a semantic network for Example 1 might look like this:

Spreading Activation

An algorithm for making connections using a semantic network. Also known as marker passing, intersection search.

Basic algorithm:

1. “Get rope” node activated by first sentence.

2. “Commit suicide” node activated by second sentence.

3. Activatioin spreads in both directions.

4. As soon as they intersect, you have one possible explanation.

Problems with spreading activation

“Get a rope” could be useful for other things – e.g. drying clothes. Activation could take an undesired path and potentially keep spreading forever without intersection.

Solution: Introduce a decay factor, say b = 0.8. The first node is given activation level a, the second ab, the third ab², and so on. So the activation becomes less significant as spreading goes on.

Suppose multiple paths lead to the same intersection. For instance, you could use a rope to scale a building so you could commit suicide by jumping off. One could accept the shortest path, but 1.) paths may be the same length and 2.) not all relationships between nodes are equally likely.

Solution: Add a “link strength” to the links.

Suppose we add another sentence, and then shift our focus to look at the 2^nd and 3^rd sentences (assume we have a working memory of 2 sentences). What happens to the activations made by the 1^st sentence? They should stay but be of decreasing importance as time goes on.

Solution: Add another decay factor, say g = 0.7. Every time we move to the next sentence, we set each activation level = g * (activation level). Over time, old activations will fade.

Overlay of Semantic Networks

Recall the question of example 2: “Does AB look more like C1 or C2?” To help answer this, we could overlay the semantic network representation of AB with another semantic network (shown in red) that captures the structure of the first network:

We can construct similar overlays for C1 and C2, and then compare the overlays to see which sets of diagrams are most similar.

Roles of Semantic Networks

1. Given a library of situations described as an overlay of semantic networks, can decide which is closest to the input.

2. Can relate different parts of input.