Related Experiment Videos
Pipelining of Fuzzy ARTMAP without matchtracking: correctness, performance bound, and Beowulf evaluation
José Castro1, Jimmy Secretan, Michael Georgiopoulos
1Department of Computer Engineering, Technological Institute of Costa Rica, Cartago, Costa Rica. jcastro@itcr.ac.cr
Summary
This study speeds up Fuzzy ARTMAP neural network training using a pipeline parallelization technique. This method achieves linear speedup for classification tasks, making large datasets more manageable.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Computational Neuroscience
Background:
- Fuzzy ARTMAP neural networks are effective classifiers but suffer from slow convergence with large training datasets.
- Training time increases significantly with the number of patterns, limiting practical applications.
Purpose of the Study:
- To investigate the convergence time of Fuzzy ARTMAP.
- To propose and analyze a coarse-grain parallelization technique to accelerate Fuzzy ARTMAP training.
- To evaluate the parallel implementation's performance on large datasets.
Main Methods:
- A pipeline parallelization approach was developed for Fuzzy ARTMAP, specifically excluding the match-tracking mechanism.
- Theoretical analysis and proofs were provided for the parallel implementation's characteristics.
- Experiments were conducted on a BEOWULF cluster using the Forrest CoverType dataset and two artificial Gaussian-distributed datasets.
Main Results:
- The proposed pipeline parallelization technique demonstrates linear speedup relative to the number of processors used.
- The parallel Fuzzy ARTMAP implementation effectively reduces training time on large-scale classification problems.
- Performance was validated across diverse datasets, including real-world and synthetic data with varying class overlap.
Conclusions:
- Coarse-grain pipeline parallelization is an effective strategy for accelerating Fuzzy ARTMAP training.
- The parallel approach maintains desirable classification characteristics while significantly improving computational efficiency.
- This research enables the scalable application of Fuzzy ARTMAP to complex, large-volume classification challenges.
