Related Experiment Videos
A novel information geometric approach to variable selection in MLP networks
A Eleuteri1, R Tagliaferri, L Milano
1Dipartimento di Scienze Fisiche, Università degli Studi di Napoli Federico II, via Cintia, I-80126 Napoli, Italia. eleuteri@na.infn.it
Summary
This study introduces an information geometry approach for selecting variables in multi-layer perceptron networks. The method efficiently searches and ranks network models by analyzing divergences and posterior probabilities.
Area of Science:
- Machine Learning
- Information Geometry
- Neural Networks
Background:
- Multi-layer perceptron (MLP) networks are widely used but require effective variable selection.
- Existing variable selection methods may not fully leverage the geometric properties of network models.
Purpose of the Study:
- To develop a novel information geometric-based criterion for variable selection in MLP networks.
- To enable efficient search and ranking of network models based on input dimension reduction.
Main Methods:
- Utilizing projections of Riemannian manifolds defined by MLP networks.
- Applying divergence measures between projected submanifolds to assess model differences.
- Evaluating posterior probabilities of projected models for ranking.
Main Results:
- Demonstrated the efficacy of information geometric divergence for efficient input space search.
- Successfully ranked projected models using posterior probability evaluation.
- Validated the algorithm's performance on both synthetic and real-world datasets.
Conclusions:
- The proposed information geometric-based criterion offers an effective approach for variable selection in MLP networks.
- The method provides a robust way to search and rank models, outperforming or matching existing techniques.