A Collective Neural Architecture with Local Attractors and Multi-Level Consensus

This document describes an alternative neural network architecture where hidden neurons are organized into groups that function as local collectives. Each neuron is treated as an individual agent with its own attractor target, and local consensus mechanisms ensure coordination within groups, while soft interactions align different groups. This multi-level structure maintains diversity of hidden activations and offers an alternative to traditional uniform layers. We illustrate the approach with experiments on MNIST classification, showing comparable accuracy to standard models while preserving internal specialization across neuron groups. https://huggingface.co/spaces/Dreanhunter30/AI_encoder_decoder_VAE_MOE ( over here you could find alreafy trained agent on this architecture. This minimal mvp project) Автор: Бато Найданов Лицензия: CC BY-NC 4.0.