The interactive activation model (IAM), introduced by McClelland and Rumelhart (1981), is one of the most influential connectionist models in cognitive science. Designed to explain visual word recognition, it demonstrated how processing could proceed simultaneously at multiple levels — features, letters, and words — with information flowing both bottom-up and top-down through excitatory and inhibitory connections. The model provided a mechanistic account of the word superiority effect: the finding that letters are identified more accurately when embedded in words than when presented in isolation or in nonwords.
Architecture
Between levels: excitatory connections (consistent units) and
inhibitory connections (inconsistent units)
Within levels: mutual inhibition (lateral inhibition)
Activation update:
If net input > 0: Δaᵢ = net · (max − aᵢ) − decay · aᵢ
If net input ≤ 0: Δaᵢ = net · (aᵢ − min) − decay · aᵢ
The model consists of three processing levels. The feature level contains detectors for visual features (horizontal, vertical, and diagonal line segments) at each letter position. The letter level contains detectors for each letter at each position. The word level contains a unit for each word in the model's vocabulary. Connections between consistent units at adjacent levels are excitatory (a horizontal line in position 1 excites the letter "A" at position 1), while connections between inconsistent units are inhibitory. Crucially, connections are bidirectional: word-level activation feeds back to boost the activation of the word's constituent letters.
The Word Superiority Effect
The word superiority effect emerges naturally from the IAM architecture. When the word "WORK" is presented, bottom-up activation from features excites the correct letters, which in turn excite the word unit for WORK. The WORK unit then feeds excitatory activation back down to its constituent letters, boosting their activation above what they would achieve from bottom-up input alone. This top-down support enhances letter identification in words but not in nonwords (which do not have corresponding word units). The model also explains the pseudoword superiority effect: letter identification is better in pronounceable nonwords than in unpronounceable strings, because pseudowords partially activate real word units, providing some top-down support.
The IAM was a key salvo in the debate over interactive versus modular processing. Fodor's (1983) modularity thesis proposed that perceptual systems are informationally encapsulated: higher-level knowledge cannot penetrate lower-level processing. The IAM's top-down connections directly contradicted this, showing that word-level knowledge could influence letter-level processing. This debate continues in perception research, with the IAM providing one of the strongest computational demonstrations that interactive processing can explain empirical phenomena that are difficult to account for with purely bottom-up models.
The interactive activation framework has been extended to model speech perception (the TRACE model of McClelland & Elman, 1986), object recognition, and semantic processing. Its core principles — parallel processing at multiple levels, bidirectional excitation between consistent representations, lateral inhibition within levels, and the gradual settling of activation into a stable pattern — have become foundational design principles for connectionist models in mathematical psychology.