The A.I. Daily
Real fly connectome failed to learn Pong; audit reveals issues
A researcher tried to train a subgraph of the MaleCNS v1.0 fly connectome to learn Pong but observed no learning. An audit uncovered issues: a neuPrint regex bug zeroed two neuron populations, the initial neuron selection lacked paths from photoreceptors, and half the motor neurons had zero sensory synapses. Rebuilding and tracing found a different descending neuron that actually connected end-to-end, illustrating how failures revealed critical data and selection errors.