Von Neumann Networks: A New AI Architecture from a 1950s Brain Model

A recent paper on arXiv presents Von Neumann Networks (VNNs), a neural network framework inspired by John von Neumann's cellular automaton model from the mid-20th century. Developed in the 1950s, von Neumann's initial system organized cells in an array, each assigned a specific role or state, which was expected to adhere to a diffusion process. This new research demonstrates that by integrating this idea into contemporary deep learning, it is possible to create an artificial neuron with customizable, learnable roles, referred to as the Von Neumann neuron. Networks utilizing these neurons achieve a self-organized design, with architecture determined by the arrangement of inputs and outputs on the cellular grid. The mathematical foundation for VNNs involves extending neural operators and learning Green's functions through convolutions on a cellular structure. The paper can be found on arXiv under ID 2605.05780.