MC-PSO and MC-APSO: Parallel RBFNs beat standard methods on accuracy and speed
New multi-column RBFN with swarm optimization outperforms gradient-based and single-column approaches across benchmarks.
A new paper introduces two hybrid neural architectures that tackle scalability issues in radial basis function networks (RBFNs). Traditional training methods like gradient-based ErrCor and population-based PSO face excessive kernel computations and large hidden layers on big datasets. The authors propose Multi-Column RBFN with PSO (MC-PSO) and its adaptive variant MC-APSO, inspired by the success of the earlier Multi-Column RBFN (MCRN) approach. In these designs, multiple small RBFNs are trained in parallel, each on a distinct spatial subset of the data using PSO or APSO. During inference, only the RBFNs whose training regions neighbor the test sample contribute to the final output, reducing computation while preserving accuracy.
The experimental evaluation on several benchmark datasets shows clear gains: MC-PSO and MC-APSO outperform ErrCor, standard PSO, APSO, and MCRN in both accuracy and recall. Additionally, the models demonstrate faster training and testing times in most experiments. The parallel structure and specialized local training enable the swarm optimization to escape local minima more effectively and scale to larger datasets. This work highlights how combining evolutionary algorithms with modular neural architectures can improve performance and efficiency in machine learning tasks.
- MC-PSO and MC-APSO train multiple RBFNs in parallel, each on a spatial subset of the dataset using PSO or adaptive PSO.
- Outperform ErrCor, PSO, APSO, and prior MCRN on benchmark accuracy and recall metrics.
- Achieve faster training and testing times while reducing kernel computation overhead.
Why It Matters
Enables more scalable and accurate neural network training for large datasets, combining swarm intelligence with parallel architecture.