The approach adds a second layer of modulation inside neural nets, enabling neuron-to-neuron influence across the same layer, which we model mathematically and implement efficiently. We call these Ephaptically Coupled Artificial Neural Networks (EC-ANNs).
Preliminary testing shows strong cross-domain gains compared to standard weight-only baselines: a 74.65% reduction in perplexity in language (GPT-2, WikiText-103), a 3.08% accuracy gain in vision (ResNet-18, ImageNet-100), and a 6.67% reward improvement in reinforcement learning (PPO, Walker2d-v5). These performance improvements were achieved with minimal parameter overhead using pretrained models: ~0.5% for GPT-2 small, ~2.2% for ResNet-18, and ~0.98% for the PPO network. Initial ablation studies suggest that ephaptic coupling provides a highly efficient means of architectural enhancement compared to conventional width expansion.
We're currently applying this to AI-based predictive maintenance for 5G infrastructure at my company, Bloxtel.
Would love feedback, questions, or challenges. This idea has been on my mind for years, and I finally went public after revisiting a 1969 interview with McCulloch (of McCulloch-Pitts neuron fame). His words about discovery inspired me to share.
Paper (may require cookies): http://bloxtel.ai