MultiBeeSketch: A Pure Integer Neuromorphic Architecture for Real-Time Streaming Predictions

This paper introduces MultiBeeSketch, a hardware-friendly neuromorphic prediction architecture that operates entirely within the pure integer domain. Standard predictive models rely on floating-point arithmetic, massive matrix multiplications, and gradient descent, rendering them inefficient for ultra-low-power edge computing. MultiBeeSketch replaces weight-based neural layers with sparse count boards (BeeBoards) and independent, multi-scale hash-indexed predicting workers. By employing distinct pseudo-random seeds and a voting consensus system, the architecture eliminates the risk of memory collisions and achieves a context depth of up to 16 symbols with an incredibly small memory footprint.