Adaptive Mixtures of Local Experts

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Neuroscience

Background:

• Supervised learning typically involves single, complex networks.
• Existing methods like multilayer networks and competitive learning have distinct mechanisms.
• A need exists for integrated approaches that leverage modularity.

Purpose of the Study:

• Introduce a new supervised learning procedure for systems with multiple separate networks.
• Establish a connection between modular supervised networks and associative competitive learning.
• Demonstrate the efficacy of this procedure in task decomposition.

Main Methods:

• Developed a supervised learning procedure for distributed network systems.
• Each network is trained on a subset of the data.
• The procedure integrates aspects of multilayer and competitive learning paradigms.

Main Results:

• The proposed learning procedure effectively divides complex tasks into simpler subtasks.
• Demonstrated successful application to a vowel discrimination task.
• Showcased that each subtask can be solved by a simple expert network.

Conclusions:

• The new procedure offers a modular approach to supervised learning.
• It bridges the gap between multilayer and competitive learning strategies.
• This method enables efficient task specialization within a network system.