M🍄Tree

I trained the fly to recognize the US flag, salute, and buzz the Pledge of Allegiance in Morse code. It works by injecting image pixels into a rate-based model built from FlyWire’s 139,255-neuron connectome (~15 million directed connections). A trained readout classifies the resulting network activity, triggering a learned Morse sequence and an optimized foreleg salute in a MuJoCo fly body. I trained three separate components around a fixed FlyWire connectome. Recognizing the flag: I projected flag-image pixels into 4,096 artificially selected input neurons, propagated activity through the full network for 12 rate updates, and extracted 128 downstream activity features. A small 128→64→32→1 classifier learned “American flag → respond; other flags → silence,” using AdamW and weighted binary cross-entropy for 1,200 epochs. Training included 256 flag designs, image perturbations, and blank controls. It correctly classified 550/550 held-out examples, consisting of new mild transformations and controls. Learning the pledge sequence: I encoded the pledge into Morse symbols and trained a 192-unit GRU to predict each next symbol. During training, it received the correct preceding symbol; during evaluation, it had to generate the sequence from its own previous outputs. At epoch 120, it reproduced all 492 tokens exactly, including the stop token. Engineered timing converts those symbols into short and long bursts of a 200 Hz buzz. This learns one prescribed sequence; it does not learn language. Finding the salute: I optimized seven foreleg joint targets inside MuJoCo. A cross-entropy search evaluated 1,152 candidate postures, retaining better candidates and sampling around them, followed by numerical refinement. The objective minimized distance between the forefoot and a target near the forehead. Across eight perturbed starting states, mean error fell from approximately 1.58 mm to 0.019 mm.