connectionism1

Computational Theory of Mind

• Information processing involves symbol manipulation via rules.

• Distinction made between:

  • Structure (symbols)

  • Process (rules, algorithms)

• Types of processing:

  • Sentential

  • Propositional

  • Sequential

• Language of thought concept.

Connectionist Theory of Mind

• Alternative view of information processing, also known as:

  • Parallel Distributed Processing (PDP)

  • Neural Networks

• Historical contributions:

  • McCulloch & Pitts (1940s)

  • Rosenblatt (1960s) - perceptron (2-layer network)

Applications of Connectionist Networks

• Commonly utilized to classify patterns by defining decision regions in pattern space.

• Examples from Churchland’s "The Engine of Reason, the Seat of the Soul" include:

  • Color Space: Classification of colors based on opponent cells.

  • Taste Space: Categorization of tastes with respect to different activation levels for various substances.

  • Face Space: Parameters like nose width, eye separation, and mouth fullness classified in three dimensions.

Connectionist Networks Overview

• Definition: Networks of simple units (processors) operating simultaneously (in parallel) and inspired by biological processes.

Processing Stages

1. Input Phase: Input information analogous to perception.

2. Computation Phase: Transformation of input analogous to thinking.

3. Output Phase: Generates output analogous to action.

Structure of Connectionist Networks

• Each processing unit behaves as a simplified neuron:

  • Computes total incoming signals from previous layers.

  • Adopts level of internal activation based on these signals.

  • Generates outgoing signals modulated by connection weights.

Connection Weights

• Function as communication channels between processing units (similar to synapses).

• Can be:

  • Inhibitory (negative weights)

  • Excitatory (positive weights)

  • Strength of connection varies based on absolute values (range: -1 to +1).

Training Connectionist Networks

• Initial connection weights assigned randomly.

• Network trained via presenting patterns associated with known responses.

• Error term calculated as the difference between desired and actual outputs:

  • Used to adjust connection weights via backpropagation.

  • Trained until error cannot be further reduced.