Introduction
Connectionism is an approach to modeling cognition using networks of simple interconnected processing units. The activity of these units and the strengths of their connections can produce complex patterns of information processing. Connectionist models are commonly associated with artificial neural networks, although computational models are simplified representations rather than complete copies of biological brains.
Central Idea
Cognitive abilities can emerge from the interactions of many interconnected units. Learning can occur when the strengths of connections are adjusted in response to experience, examples or feedback.
Key Components
| Component | Explanation |
|---|---|
Processing Units | Computational elements that receive information, combine inputs and produce activation. |
Connections | Links through which activation or information passes between units. |
Connection Weights | Numerical values representing the strength and influence of connections. |
Activation | The current level of activity of a unit, calculated from its inputs and an activation function. |
Learning Rule | A procedure for changing connection weights or other model parameters based on experience or feedback. |
Network Architecture | The arrangement of units and connections within the model. |
Network Organization
| Type | Explanation |
|---|---|
Feedforward Network | Information flows through the network in a directed sequence from input units toward output units. |
Recurrent Network | Connections allow information or activation to influence processing across time through feedback or recurrent pathways. |
Distributed Representation | Information is represented by a pattern of activity across multiple units rather than necessarily by one dedicated unit. |
Learning Process
Provide the network with input information.
Allow activity to spread through connected units.
Generate an output or prediction.
Compare the output with a target or other learning signal when the training method uses one.
Adjust connection strengths according to the learning procedure.
Repeat across training examples and evaluate performance on new data.
Types Of Learning
| Type | Explanation |
|---|---|
Supervised Learning | The model learns from examples that include target outputs. |
Unsupervised Learning | The model identifies patterns or structure in data without supplied target labels. |
Reinforcement Learning | The model learns through interaction and feedback in the form of rewards or penalties. |
Contributions To Cognitive Science
Provides models of learning through changes in connection strengths.
Offers ways to investigate distributed representations.
Helps researchers study pattern recognition and generalization.
Supports theories of how cognitive abilities may emerge from many interacting processes.
Provides computational approaches for comparing cognitive behavior with artificial network performance.
Limitations And Debates
A successful artificial neural network does not automatically demonstrate that human cognition uses the same mechanism.
Some models require large datasets and substantial computational resources.
Distributed representations may be difficult to interpret.
Models can learn unwanted statistical patterns from biased or incomplete data.
Different network architectures and learning rules may produce substantially different behavior.
Misconceptions
Connectionist models are inspired by some properties of neural systems but are not exact simulations of the brain.
Not all connectionist models have the same architecture or learning mechanism.
Connectionism does not mean that cognition has no structure or representation.
The term connectionism is broader than any single modern machine-learning architecture.
Knowledge Check
| Question | Answer |
|---|---|
What changes during learning in many connectionist models? | Connection weights or other model parameters are adjusted based on experience or learning signals. |
What is a distributed representation? | A representation in which information is carried by a pattern of activity across multiple connected units. |