Seq103 neuroevolution framework cuts parameters 160,601x with minimal accuracy loss

A new neuroevolution framework called Seq103 promises to dramatically shrink sequence models without sacrificing much accuracy. Developed by Wenxiao Li, Yongjian Liu, and Qing Xie, Seq103 applies evolutionary algorithms to jointly discover network topology and weights. It builds on NEAT (NeuroEvolution of Augmenting Topologies) with a shared backbone that uses node-and-connection representation, per-class RMSE evaluation, mutation-based evolution with class-wise recombination, and elitism. An optional recurrent extension adds hidden-state nodes and connections for step-wise temporal memory, enabling the same search pipeline to work on both feedforward and recurrent tasks.

Tested on 8 text datasets and the entire 128-dataset UCRArchive2018 time-series benchmark, Seq103 achieved remarkable parameter efficiency. On step-wise recurrent tasks, it retained 86.96% of the best baseline accuracy while using 34.6x to 3,218x fewer parameters. On sample-wise feedforward tasks, it retained 81.95% accuracy with a parameter reduction between 11.8x and a staggering 160,601x. These results suggest Seq103 can produce ultra-compact models suitable for edge deployment or memory-constrained environments, though accuracy trade-offs exist. The paper is available on arXiv (2606.07664).