Von Economo neurons key to reliable learning in spiking neural networks

A new computational study published on arXiv by Esila Keskin provides a formal account of how Von Economo neurons (VENs) enable consistent learning in recurrent spiking neural networks. The VENCircuit model embeds 40 VEN-like projection neurons (2% of total) into a pyramidal circuit, trained on a binary classification task. Across 50 random initializations, VEN-intact networks converged in 98% of cases (49/50) versus 70% (35/50) without VENs (odds ratio 21.0, p=8.7e-5). Crucially, failed VEN-ablated networks showed complete absence of learning—not merely slower learning—indicating VENs are essential for reliable convergence.

The study reveals VENs function as direct gradient pathways that bypass Jacobian instabilities in the recurrent circuit. Phase-ablation experiments showed that removing VENs during mid-training (epochs 5–25) was most disruptive, as the pyramidal circuit develops a co-adaptive dependency. Inference-time VEN ablation caused significant performance drops (p=0.022), ranging from no change (16/20 networks) to catastrophic collapse from 0.989 to 0.620 accuracy. This work provides falsifiable predictions for organoid and electrophysiology studies, linking VEN loss in behavioral-variant frontotemporal dementia (bvFTD) and reductions in autism spectrum conditions (ASC) to stochastic failures in social skill acquisition.