Connectionism: Introduction

The Basics

Connectionism, neural networks
- The idea that the hardware matters
- Neurally-inspired cognitive modeling
What is a connectionist model?
- A network
  - A vector of activations (length n), corresponding roughly to the firing patterns of neurons
  - Sometimes a separate vector of outputs
  - A matrix of weights (dimensions n X n), corresponding roughly to the structures that connect neurons to one another: axons, dendrites; synapses
- A task
  - A set of input vectors
  - (Sometimes) a set of target vectors
- An activation rule
- A weight (learning) rule
A connectionist unit (corresponding roughly to a neuron or a cluster of neurons)

Input and output
- Matrix-memory models: all units participate potentially in input and output
- Others: separate input, output, and sometimes hidden units
Representations
- Localist
- Distributed: input, output, hidden
  - Multiple units participate in the representation of each concept.
  - Multiple concepts are represented by each unit.
Continuous and discrete time
Static and sequential networks
Learning
- Supervised
- Reinforcement
- Unsupervised
Connectivity
- Feedforward
- Partially recurrent
- Completely recurrent; constraint satisfaction

Learning: Basic Concepts

Connectionism: Connectionist vs. Symbolic Models

Last updated: 26 February 1996
URL: http://www.indiana.edu/~gasser/Q351/connectionism1.html
Comments: gasser@salsa.indiana.edu
Copyright 1996, The Trustees of Indiana University