AI Glossary / Deep learning / Intermediate

Interactive activation and competition networks

Definition

Interactive activation and competition networks, or IAC networks, are recurrent connectionist models in which mutually supportive representations excite one another while competing alternatives inhibit one another. Units are commonly organized into pools of alternatives, with supportive connections across pools and inhibitory connections within a pool. The evolving activity pattern can model context-sensitive recognition, memory retrieval, and pattern completion.

What it means

Evidence need not move only from raw features toward a final decision. Higher-level and lower-level representations can influence each other through reciprocal connections. A partial cue may activate several alternatives, which then compete while receiving additional support from related features. The resulting interpretation depends on both the cue and the network's stored associations. IAC is a family of cognitive modeling mechanisms, not a synonym for an interactive chatbot or a general guarantee of accurate memory.

How it works

At each step, a unit receives a weighted sum of other units' outputs plus any external input. Positive connections supply excitation and negative connections supply inhibition. An activation-update rule combines that net input with current activity, bounds, and decay toward a resting level. In a common formulation, only positive activation contributes transmitted output. Repeated updates allow feedback and competition to reshape the pattern. The weights may be specified to encode a knowledge structure in a demonstration; the activation dynamics should not automatically be mistaken for a learning algorithm that changes those weights.

Three practical examples

• Word-recognition modeling: letter and word representations interact so that context can influence interpretation of ambiguous visual evidence.
• Associative retrieval: activating some known properties helps retrieve other properties associated with stored individual examples.
• Cognitive experiments: researchers examine priming, competition, or persistence by changing initial activity, external cues, and connection strengths.

Think of it this way

Think of several possible interpretations receiving support from compatible clues while competing with mutually exclusive alternatives. Agreement among clues can reinforce one interpretation. The analogy is limited because the network follows numerical activity rules, and mutual reinforcement can also sustain an incorrect interpretation if the connections or cues favor it.

Worked example

Consider two competing candidate units, A and B, in the same pool. A cue unit has positive output 0.8 and connects to A with weight 0.5 and to B with weight 0.25. Suppose A and B currently have positive outputs 0.6 and 0.4, and their mutual inhibitory weight is -0.3. Ignore other inputs for this calculation. A's net input is 0.5 times 0.8 minus 0.3 times 0.4 = 0.28. B's is 0.25 times 0.8 minus 0.3 times 0.6 = 0.02.
Both candidates receive support from the cue, but A receives more and suppresses B more strongly. These numbers are net inputs, not the next activation values: the specified bounded update and decay rule still have to be applied. If the cue changes, the later trajectory can change too. Adding reciprocal links from candidates to cue-related representations would introduce the cross-pool feedback characteristic of a fuller IAC model. This snapshot illustrates competition without claiming that A must eventually win under every parameter setting or starting state.

Illustrative case study

A fictional perception-research team builds an IAC demonstration for recognizing ambiguous symbols in short words. Strong word-context input sometimes favors a letter inconsistent with the visual stimulus. Rather than hiding these cases, the team uses them to examine the balance between external evidence and recurrent support. It compares predicted response patterns with the intended experiment and distinguishes a model of contextual influence from a system guaranteed to read every symbol correctly.

Common misconceptions and mistakes

Interactive activation does not mean that all units excite each other; the distinction between supportive and competitive connections is central. Activation is not automatically a calibrated probability. A model settling into a state does not prove that the interpretation is true. Feedback can create persistence and dependence on prior state, so initialization and cue duration matter. Finally, a hand-configured IAC network can perform retrieval dynamics without learning its connections from data, which differs from the training process of many modern neural models.

Business use

For business readers, IAC is most useful for understanding how context and competition can influence recognition or retrieval. If adapted to a practical decision system, its associations and update rules need explicit justification and evaluation. Test contradictory cues and ambiguous inputs, not only easy examples. Explain whether displayed activations are internal support values or validated probabilities. The method's cognitive inspiration can motivate experiments, but successful simulation of one psychological pattern does not establish broad operational reliability.

People also ask

What makes activation interactive?

Reciprocal connections allow activity in one pool to affect another pool and receive influence back.

What creates competition?

Inhibitory connections among alternative units reduce their activity relative to more strongly supported competitors.

Does the model necessarily learn while it settles?

No. Settling changes unit activations; changing connection weights requires a separately specified learning mechanism.

Check your understanding

A unit has net input 0.3 in an IAC calculation. Can its next activation be stated as 0.3 without further information?

Answer: No. Net input must be combined with its current activation, decay, bounds, and the particular update rule.

Related terms

• Recurrent neural network
• Associative memory
• Connectionism
• excitation
• inhibition
• pattern completion
• priming
• competitive pool

Learn next

trace a small IAC network with two candidate units and several clues, separating transmitted outputs, net inputs, and activation updates. Compare behavior under strong, weak, and contradictory cues, and observe how initialization and feedback affect the trajectory.

More in Deep learning

• Infomax - Infomax is an information-theoretic learning principle that seeks representations retaining as much information about their inputs as possible under...
• Input layer - An input layer is the entry interface through which a neural network receives its data representation.
• Instantaneously trained neural networks - Instantaneously trained neural networks are a family of feedforward models whose parameters can be assigned directly from training examples rather than...
• Kohonen Network - A Kohonen network commonly refers to a self-organizing map, or SOM: an unsupervised neural model that learns representative vectors arranged on a predefined...
• Large memory storage and retrieval neural network - A large memory storage and retrieval neural network usually refers to LAMSTAR, a historical neural architecture that combines modules based on...
• Large width limits of neural networks - Large width limits of neural networks are mathematical descriptions of what happens when the number of units or channels in selected layers grows without...

Sources

McClelland and collaborators, Interactive Activation and Competition, PDP handbook: https://web.stanford.edu/group/pdplab/pdphandbook/handbookch3 Original author-hosted chapter: https://stanford.edu/~jlmcc/papers/PDP/McClelland_EPDP/Chap2_EPDP88.pdf

Previous: Inter-relational constraint Next: Interactive evolutionary computation