1. Associative Theories of LTM

2. Networks How is all the information in our LTM represented and how does one go about finding and retrieving a bit of knowledge among the hugh store of material available in LTM? Representing information in LTM as a network of connections has been around for centuries. The following desribes the basic ideas of a “network”:

3. Networks (con’t) Nodes – represent individual ideas (e.g., a “South Dakota” node). Association links – connect nodes to other nodes. A “search,” then, consists of travelling from one node to the next along the links until the target information is reached.

4. Spreading Activation A node is activated when it receives sufficient input or excitation. At that point, activation spreads along its links to other nodes, partially activating those nodes. Nodes receive activation from other nodes, increasing “subthreshold activation” levels. when the “sum” of activation reaches “response threshold,” the node “fires.” Note the parallels between our conceptualization of activation in a network and the way neurons work.

5. Psychological Evidence Of Networks The results of a variety of lines of research are consistent witht the idea of a network and predictions based on a network: 1.Cueing - When recall alone is insufficient to activate a node, a “cue” or “hint” can sometimes help by activating another node that can spread its activation to the target node. 2.Context Reinstatement - When study and test contexts are the same, recall is better because the nodes that were activated and connected to the new material will likely be activated again during testing.

6. Evidence Of Networks (con’t) 3.Lexical-Decision Task (priming) - In a variation of the lexical decision task, pairs of words or nonwords are presented. Some of the word pairs were related (e.g., bread- butter) while others were not (e.g., nurse-butter). Spreading activation resulting from the “bread” node should partially activate the “butter” node, allowing for a quicker response time.

7. Evidence Of Networks (con’t) 4.Sentence Verification - Since one travels from node-to-node, it should take longer to reach a more distant node than a closer node. Consider the following partial network: When asked to verify the truth of statements such as, “A robin is a bird” and “A robin is an animal,” the first sentence should be verified more quickly than the second.

8. Evidence Of Networks (con’t) In general, the predictions are confirmed. From this research, three general conclusions can be drawn: a.If a fact about a concept is frequently encountered, it will be stored with that concept even if it could be inferred from a more superordinate concept. b.The more frequently encountered a fact about a concept is, the more strongly that fact will be associated with the concept. And the more strongly associated facts are with concepts, the more rapidly they are verified. c.Verifying facts that are not directly stored with a concept but that must be inferred takes a relatively long time.

9. The Fan Effect Some nodes have many connections (e.g., robin) while others have few (e.g., aardvark). It is assumed that when a node is activated, activation will spread through all its association links (i.e., connections).

10. The Fan Effect (con’t) There is, however, a limit on the amount of activation that can spread from such a node. The more connections there are from a node, the less activation there will be spread to any one associated node. Fan Effect - the name given to the increase in RT related to the increase in the number of connections (i.e., greater fan) associated with a node.

11. A Methodological Illustration Ss learn sentences of the form “The is in the.” The number of times the person or the place occurred in the sentences varied. For example, The doctor is in the bank. (1 -1) The fireman is in the park. (1 - 2) The lawyer is in the church. (2 - 1) The lawyer is in the park. (2 - 2) Ss were given a speed-recognition test in which they were presented sentences (some previously learned and some new foils) and had to indicate whether or not they had studied the sentences.

12. A Methodological Illustration (con’t) Ss were fastest when responding to (1 - 1) type sentences and slowest when responding to (2 - 2) type sentences: they were equally fast at responding to (1 - 2) and (2 - 1) type sentences.

13. Searching A Network How might we search through such a network as we’ve described? A search on the Internet can be used as an analogy. One advantage our network model, however, is that activation can spread from more than one source simultaneously, resulting in the convergence of activation to the sought-after node. How do we get to an “entry node”? Our networks rely, in part, on sensory input. Feature nodes and spreading activation of the feature nets are seamlessly connected to our memory network and, therefore, activate the cognitive nodes we have been discussing.

14. What’s In A Node? How is information (in general) represented and, subsequently, how is complex information represented? Several models have been proposed: Node = Concept Propositional Networks

15. Node = Concept This model proposes each node represents a concept (e.g., “Lincoln,” “war,” etc). Each node is connected via a “relational” associative link (e.g., “isa” or “hasa”). Such a conception, however, would require too many types of relational links (e.g., opposite of, analogous to, larger than, etc) to be useful.

16. Propositional Networks A more fruitful proposal comes from “propositional” networks, most notably Anderson’s ACT (Adaptive Control of Thought) computer program. A “proposition” is the smallest unit of knowledge about which it makes sense to judge as being true or false. Consider the following sentence: “Lincoln, who was president of the USA during a bitter war, freed the slaves.”

17. Propositional Networks (con’t) That sentence is made up of several simpler sentences, each representing a proposition: A:Lincoln was president of the USA during a war. B:The war was bitter. C:Lincoln freed the slaves. Note each of those simpler sentences can be judged true or false. We can represent that information in a propositional network:

18. Propositional Networks (con’t) Each proposition is represented by an ellipse with labeled arrows to its relation and arguments. The ellipse, relations, and arguments are called “nodes” and the arrows are called “links” (i.e., they connect the nodes). Lincoln president-of USA war time relation agent object war subject relation bitter agent Lincoln object slaves relation freed The nodes can be thought of as “ideas” and the links as “associations” between the ideas and identified by their syntactic role within the proposition.

19. Propositional Networks (con’t) Once the nodes are all connected, they are organized within the network for ease of interpretation.

20. Propositional Networks (con’t) In addition to representing simple ideas as the sentence above, propositional networks can represent more complex ideas or knowledge. For example… This network, represents a small portion of our knowledge about “dogs.”

21. Propositional Networks (con’t) There are two types of nodes: type – refers to a general category (e.g., dog) and are true for the entire category. token – refers to a specific instance of the category (e.g., “my dog”). The token nodes are connected to type nodes:

22. Propositional Networks (con’t) “Time” and “location” nodes are also incorporated into propositions Like network models in general, nodes are connected by associative links, some of the links are stronger than others (depending on frequency and recency of use), and spreading activation partially activates connected nodes.

23. Problems With Propositional Networks Network models, though promising in their approach and having the support of many researchers, do have their limitations. Let’s examine some of those limitations: Retrieval blocks – There is a saying that “’close’ only counts in horseshoes.” That saying should also apply to the concept of “spreading activation,” yet we see many cases of retrieval blocks where it seems “close” doesn’t count (e.g., TOT phenomenon). Too many distant connections – Suppose you activate a node that has many links connected to it (e.g., “health”). Spreading activation will result in all those associated nodes to be activated as well. If the information you seek it one or two additional links away, you will have activated tens of thousands of other nodes. How do you sift through them all?

24. Addressing Those Challenges Given that spreading activation will activate nodes you seek and many that are irrelevant, a means of narrowing activation would be useful, allowing you to focus on just the relevant nodes. That can be accomplished by postulating that nodes can inhibit as well as excite adjacent nodes via spreading activation. A more strongly activated node would have the effect of deactivating neighboring, less active nodes, until only one node remains active. This is referred to as a winner-takes-all system. Such a system could address the problem of activating too many nodes and narrowing the focus of the search for a piece of information.

25. fGULEIWSQHTAXfGULEIWSQHTAX BPNCJMFZDKMORVYBPNCJMFZDKMORVY fGULEIWSQHTAXfGULEIWSQHTAX BPNCJMFZDKMORVYBPNCJMFZDKMORVY fGULEIWSQHTAXfGULEIWSQHTAX BPNCJMFZDKMORVYBPNCJMFZDKMORVY BPNCJMFZDKMORVYBPNCJMFZDKMORVY Connectionism “Connectionist networks” eliminate the idea of a node representing an individual idea. Instead, ideas are considered to be patterns of activation across the network… distributed representations. Here is a simple illustration:

26. Connectionism (con’t) In this conception, an individual node has no particular meaning. Instead, the entire pattern of nodes must be considered to determine what is being represented... its meaning is “distributed” across the entire network. Pattern activation is accomplished quickly through parallel distributed processing (PDP) and without the aid of a “central executive.” One advantage of the connectionist network is that it can perform simultaneous multiple constraint satisfaction. That is, given a problem with several constraints, different parts of the network can work independently on each constraint and collectively come up with an average-like solution to the problem.

27. Learning In A Connectionist Network To say that you “know” something means that you have an available pattern of nodes that represents that knowledge. But how did that pattern emerge in the first place? How does a connectionist network “learn?” Connection weights – the strength of individual connections – are adjusted locally by on-going activity. The adjustments are accomplished by algorithms: “What-goes-with-what” “Feedback”

28. Current Status of Connectionist Network Many researchers are excited about the potential of connectionist network, in part because of a number of successes: Others, however, are not so convinced. Learning is slow and occurs only when stimuli are presented in the correct order. learned to generalize simple shapes learned to read learned to play strategic games (e.g., chess, backgammon) seem to fit well with our understanding of the functioning of neurons and the brain