NeuronSoup: Evolving Asynchronous, Shared-Neuron Temporal Graphs without Backpropagation

What the math gives to ML

The paper introduces an event-driven computation graph in which hidden neurons are shared by many input-output paths and retain an accumulator that is modified by each arriving signal. The transferable asset is the combination of asynchronous delayed propagation, state-dependent interference, and physical parameter sharing across paths. This creates a sparse recurrent architecture whose effective depth and interaction pattern vary with the input and whose computation can be executed with an event queue rather than dense layer synchronization. A practical first transfer is to replace a small MLP block with a sparse shared-neuron graph, initially trained by backpropagation through the event sequence, and compare accuracy, parameter count, activation memory, and measured event cost against a dense MLP.