HSNNM: A Hebbian Sparse Neural Network Model for Efficient Transformer Learning

Author: Kastiel Tjuandra

DOI: 10.5281/zenodo.21271668

GITHUB: The link can be found on my Zenodo page.

Abstract

We introduce the Hebbian Sparse Neural Network Model (HSNNM), a biologically-inspired transformer architecture that replaces the standard dense Feed-Forward Network (FFN) with a dynamically routed system of sparse, competing expert sub-networks termed "lobes". HSNNM integrates three core mechanisms: k-Winners-Take-All (k-WTA) lateral inhibition, non-gradient Hebbian plasticity, and net2net structural neurogenesis. We demonstrate that HSNNM achieves a 49.6% reduction in FLOPs per token (7.18M vs 14.25M FLOPs/token) compared to a matched dense baseline, while incurring a 3.4 percentage point accuracy trade-off (76.06% vs 79.45%), a favorable efficiency-accuracy balance given the significant compute savings. Through ablation studies, we show that both Hebbian plasticity and k-WTA sparsity contribute to stability over medium term training (5,000 steps). Our results establish HSNNM as an efficient and robust alternative to dense transformers.

Download

Download PDF

Citation

@misc{tjuandra2026hsnnm,
  title={HSNNM: A Hebbian Sparse Neural Network Model for Efficient Transformer Learning},
  author={Tjuandra, Kastiel},
  year={2026},
  publisher={Zenodo},
  doi={10.5281/zenodo.21271668}
}