[ CogSci Summaries home | UP | email ]

Anderson, J. R. (1983). A spreading activation theory of memory. Journal of Verbal Learning and Verbal Behavior, 22, 261-295

  author = 	 {John R. Anderson},
  title = 	 {"A spreading activation theory of memory"},
  journal = 	 {Journal of Verbal Learning and Verbal Behavior},
  year = 	 {1983},
  OPTkey = 	 {},
  OPTvolume = 	 {22},
  OPTnumber = 	 {},
  OPTpages = 	 {261--295},
  OPTmonth = 	 {},
  OPTnote = 	 {},
  OPTannote = 	 {}

Author of the summary: Jim Davies, 1999, jim@jimdavies.org

Cite this paper for:

ACT is a computer model of cognition. This paper focuses on the memory theory. Memory in ACT is a bunch of connected proposition-like cognitive elements. (p137) Other kinds of memory may also exist (such as temporal strings and visual images. Processing causes good encoding of a chunk into long term memory (research shows that motivation and intention are irrelevant). Repetition does help, as it is a form of processing. The more something is processed, the easier it is to retrieve. This strength decays with a power function (not an exponential function) over time (p138). The only difference between being in working or long term memory is the strength. Activation spreads to related chunks. Decay and spread has been found with both semantic and episodic memory.

Retrieval time behaves as an exponential with rate parameter of Activation. There is a bound to the amount of activation that can be pumped into a network by a source. This guarantees that the network will move to a stable asymptotic patterm (p139). Nodes can activate back as well (reverberation). Processing time in ACT is based on asymptotic level of activation rather than spreading time. This distinguishes it from many other spreading activation models.

ACT can successfully model proactive interference (p141). Proactive interference is when you learn to associate A with B, but then have trouble learning to associate A with C. This effect is minimized however if the second part is learned in a new context (different room, e.g.).

ACT also models the fan effect. This is when you learn to associate A with many other things, and so the activation is distributed and it makes it harder to retrieve any one of them. People have true facts stored in memory. When verifying a fact, they try to find it, and if they fail (enough time has passed) then they assume they do not know it. Thus what you don't know is not explicitly represented (p142).

Facts are stored thematically. A circus node will link to all circus facts. If all foils are non-circus, the fan effect disappears. If the foils are circusy, then the fan effect re-appears.

Subnodes hold groups of groups of facts. Subnodes are selected, which re-distribute activation. (p145) This is called subnode refocusing. Subjects may create experimental subnodes to see if they help. This protects from interference and prior associations.

The fan effect diminishes with practice but never disappears (p146). Reaction time decreaces as a power function.

Recognition is easier then recall because there is more of the trace presented. (p147)

Deep processing means building different activation paths between concepts. The result is that one chunk activates another reliably. Called Elaborative processing (p149)

Summary author's notes:

Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:
JimDavies (jim@jimdavies.org)
Last modified: Tue Sep 14 09:42:41 EDT 1999