Fly Servitor: a computer built from learning fly brains

Preliminary research preprint — Version 1 This preprint documents the initial development of Fly Servitor, a proposed computational architecture that uses simulated fruit fly mushroom bodies as programmable logic components. This version reports single-component logic-gate learning experiments using FlyWire connectome data and dopamine-gated learning rules. Experiments with multi-component circuits, arithmetic, memory and simple programs are ongoing; updated results will be released in later versions. This version is released early to share the concept, methods and preliminary findings. Abstract. We introduce Fly Servitor, a computer whose components are learning fly brains, and ask whether a whole learning brain region can serve as a reusable, composable component of a digital computer. Each component is a simulated mushroom body (MB) of the adult fruit fly, built from the FlyWire connectome (right hemisphere: 344 projection neurons, 2,597 Kenyon cells, 48 output neurons). Inputs are odor patterns, the output bit is the model's approach (1) or avoidance (0), and the component is programmed only by reward and punishment acting through dopamine-gated plasticity at Kenyon cell to output neuron synapses. With a learning rule that combines reward prediction error, bidirectional plasticity and protection of Kenyon cells that already carry the opposite memory, single simulated MBs learned all eight tested gates (NOT, AND, OR, NAND, NOR, XOR, XNOR, 3-input majority) in 20 of 20 individuals each, within 4 to 14 presentations of the truth table, with at most 0.6% errors per answer under input noise (at most 0.1% when three presentations are averaged). The classic dopamine-depression rule failed on up to 70% of individuals (OR 45%, NAND 60%, majority 70%). Encoding each input bit as one of two odors (dual rail) was critical: it turns XOR into a biconditional discrimination that the sparse Kenyon cell code separates readily, whereas single-rail encoding turns it into negative patterning. Because each component's output selects the odor fed to the next, signals are restored at every stage, allowing gates to be chained into adders, memory and simple programs.