Evolutionary feature selection for spiking neural network pattern classifiers

Researchers Michal Valko, Nuno C. Marques, and Marco Castelani have introduced a novel application of the JASTAP spiking neural network model for classification tasks, presenting it as a biologically realistic alternative to the standard multi-layer perceptron (MLP). The JASTAP model mimics neural spike timing more closely than traditional artificial neurons, potentially offering computational advantages for pattern recognition. The team extended an existing evolutionary procedure that simultaneously optimizes feature selection and neural network training, previously limited to MLPs, to work with JASTAP. This integration allows the system to automatically identify the most relevant input features while training the network, reducing dimensionality and computational overhead.

Preliminary experiments on the classic IRIS dataset yielded promising results: the JASTAP-based approach achieved comparable classification accuracy to larger MLPs while using significantly smaller network architectures. More importantly, the JASTAP model demonstrated robust performance on noisier data, maintaining accuracy where conventional networks degraded. This suggests spiking neural networks with evolutionary feature selection could be particularly valuable for real-world applications where sensor noise or data corruption is common. The work was published at the Portuguese Conference on Artificial Intelligence (EPIA 2005) and is now available on arXiv (2604.26654), though the paper is nearly two decades old, raising questions about its current relevance versus more recent advances in neuromorphic computing.