How do connectionist networks compute?

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Philosophy of Mind

Background:

• Connectionism is proposed as an alternative to classical computational theories of mind.
• Doubts exist regarding the computational validity of connectionist networks.
• A canonical definition of computation in cognitive science remains elusive.

Purpose of the Study:

• To demonstrate that connectionist networks possess genuine computational capabilities.
• To develop a comprehensive account of computation within connectionist models.
• To address and dispel persistent doubts about connectionism's computational credentials.

Main Methods:

• Developing a generic, explanatorily relevant account of computation.
• Analyzing the conventional understanding of connectionist computation and its limitations.
• Proposing a new framework for connectionist computation based on representational capacities of connection weights.

Main Results:

• The conventional account of connectionist computation is shown to be inadequate.
• A novel framework for understanding connectionist computation is developed.
• The representational capacities of connection weights are explained and analyzed.

Conclusions:

• Connectionist networks are demonstrably computational.
• A robust understanding of connectionist computation is achieved through analyzing connection weights.
• This work provides a foundation for reconciling connectionism with computational theories of mind.