Interactive Activation with Competition (IAC)
Biography
David E. Rumelhart earned his BA in Psychology and Mathematics from the University of South Dakota, 1963. He earned his Ph.D in mathematical psychology at Stanford in 1967. He has been a professor at University of California, San Diego, and later at Stanford. He retired in 1998 due to health reasons. In 2002 he was awarded the University of Louisville Grawemeyer Award for Psychology for his work in the field of cognitive neuroscience, along with colleague James McClelland.
Theory
Rumelhart was an early proponent of the belief that cognitive psychology needed to be developed through the approaches of formal scientific approaches, and the the development of formal theories was an integral part of this movement. His vision bridged the gap between mathematical and computer-based models of learning, and linguistic theories.
Througout his career he worked in many aspects of cognition.
Rumelhart, along with colleague James McClelland, developed an early connectionist model of word recognition in written language. They explored models to account for word context effects on letter identification. IAC neural network is an extension of competitive learning paradigm and has been used in a number of applications, such as speech perception, visual word recognition, visual perception.
The model consists of three different levels which contain processing units. Units are connected to the levels before and after them. Each connection is excitatory, meaning it makes a destination unit more active, or inhibitory, meaning it weakens activity. Each unit is also connected to each other unit within the same level by an inhibitory connection, introducing the element of competition. A network has an arrangement called an architecture.
1. Input level with visual feature units.
2. Intermediate level where units are individual letters
3. Output level where each unit is a word
Example: if the first letter in a word is "T", possible words might include "Time", "Tick", "Take", "Tart", "Top", "Table," etc. Similar words such as "Chime", "Pick", "Cake", "Part", "Fable" and so on would be inhibited. As activation is sent down to lower levels, all words beginning with "T" and words beginning with "T" will be somewhat activated.
Then suppose the second letter is "A". "Take,", "Tart,", "Table" and other words beginning with "TA" would continue to be activated while the competing effect would eliminate the others from consideration.
If the third letter is "B", "Table" and any other words beginning with "TAB" remain activated while others are competitively inhibited. Eventually the pattern of activation will settle into a stable configuration when a match is achieved, and the word is then recognized. Each iteration called a process cycle.
This model was highly influential to information processing cognitive psychologists in the 1980's. A similar model of speech recognition called TRACE was developed from the same concepts.
Learning Theory Bibliography
Harley, 1995