vencircuit-ven-gradient-scaffold

# VENCircuit: Von Economo Neurons as Gradient Scaffolds in Spiking Neural Networks ## Overview VENCircuit is a biologically motivated spiking neural network (SNN) architecture that embeds a small population of VEN-like projection neurons (K=40, 2% of total) in a recurrent pyramidal circuit. The methodology demonstrates that VENs function as **acquisition scaffolds** — they are necessary for reliably learning social-like computations, not for performing them after learning. ## Key Findings 1. **21x convergence advantage**: VEN-intact networks converge in 98% of cases vs 70% for VEN-ablated networks (Fisher's exact OR=21.0, p=8.7×10⁻⁵) 2. **Gradient pathway theory**: VENs provide a direct gradient pathway structurally immune to Jacobian product instabilities affecting the recurrent pyramidal circuit 3. **Critical gradient-flow boundary**: All networks initialize near the critical boundary (‖Wpp‖₂ ≈ 0.078, α ≈ 1.028) 4. **Mid-training dependency**: VEN removal most disruptive during epochs 5-25 when co-adaptive dependency forms 5. **Inference-time dispensability**: 80% of networks retain full performance when VENs removed at test time (Wilcoxon p=0.022) ## Core Mechanisms ### VEN-like Neuron Properties - Feedforward-only input connections - Direct output projection (bypassing recurrent dynamics) - Faster time constant than pyramidal neurons - Thick, myelinated apical dendrites morphology - Sparse local connectivity with long-range subcortical projections ### Gradient Flow Analysis (Proposition 1 & 2) - **Proposition 1**: The purely recurrent pathway operates near the vanishing/exploding gradient boundary (α ≈ 1.028 at initialization) - **Proposition 2**: VEN pathway carries gradient that is O(1) regardless of recurrent weight configuration - VENs act as structural residual connections preventing gradient attenuation ### Clinical Predictions 1. **Developmental VEN reduction (ASC)**: Impairs reliability of entering the learning process → variable social skill acquisition 2. **Adult VEN loss (bvFTD)**: Primarily affects networks that co-adapted strongly to VEN signals; most acquired representations survive 3. **Timing matters**: Social cognitive consequences of VEN loss depend critically on its developmental timing ## Methodology ### Network Architecture - Recurrent pyramidal circuit with VEN-like projection neurons - Burst-modulated Poisson spike statistics as stimulus proxy - Binary classification task - 50 matched random initializations with/without VENs ### Experimental Protocol 1. Full-training comparison: intact vs ablated across 50 seeds 2. Phase-ablation: remove VENs at different training epochs 3. Inference-time ablation: remove VENs from trained networks 4. Spectral norm analysis of recurrent weight matrices ## Activation Keywords - Von Economo neurons - spiking neural network social cognition - gradient scaffold SNN - VENCircuit - residual connections spiking - training convergence SNN ## References - Keskin, E. (2026). Von Economo neurons enable reliable social skill acquisition in recurrent spiking neural networks. arXiv:2605.17399 - He et al. (2016). Deep Residual Learning for Image Recognition - Allman et al. (2011). The von Economo neurons in frontoinsular and anterior cingulate cortex