Selecting useful groups of features in a connectionist framework.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Neuroscience

Background:

• Data often contains features from multiple sensors, with varying relevance for classification or function approximation (FA).
• Selecting optimal sensors and features is crucial for efficient and accurate model performance.
• Existing methods may not simultaneously handle sensor selection and relationship learning effectively.

Purpose of the Study:

• To propose novel connectionist schemes for simultaneous sensor selection and relationship learning.
• To develop methods capable of identifying and mitigating the impact of irrelevant or noisy features.
• To demonstrate the universal approximation property of the proposed schemes.

Main Methods:

• Development of two connectionist schemes based on Radial Basis Function (RBF) networks and Multilayered Perceptron (MLP) networks.
• Integration of sensor selection mechanisms within the network architectures.
• Utilizing input data X (subset Rp) and corresponding outputs/class labels Y (subset Rc) for training.

Main Results:

• Both RBF and MLP based schemes demonstrated the universal approximation property.
• The proposed methods successfully detected and eliminated detrimental feature groups online.
• Effective function approximation and classification were achieved even with the presence of bad features.

Conclusions:

• The developed connectionist schemes offer a robust approach to sensor selection and feature relevance identification.
• These methods enhance the reliability of classification and function approximation tasks by actively managing feature quality.
• The findings have implications for building more efficient and accurate machine learning models in data-rich environments.