every post about the mapped fly brain made it sound like they dropped a trained neural network into a game. two corrections: it is a fruit fly, not a housefly, and the map is a wiring diagram, an edge list of which neuron connects to which, with synapse counts and signs. it is a dataset. the viral demos are built on a fixed spiking circuit simulation, closer to a circuit simulator than to a deep network. it does not train. the deep learning was used to build the map, not to play the game.
the loop is cheap. a hundred and sixty six thousand neurons means a hundred and sixty six thousand float voltages, nothing for a modern machine. every millisecond, spikes add weight to target voltages, voltages leak toward rest, and past a threshold a neuron fires and goes quiet for a couple of milliseconds. input is game state encoded as currents into chosen neurons, and the pixel assignments in the famous doom example are openly made up proxies. it works anyway because the simulation cannot tell the difference.
the good part is that the code is out there and runnable. a leaky integrate and fire whole brain model in python with example notebooks, a drone example, a cube example, and a simulator that connects the wiring to a physics body. a few evenings of setup and you have your own little brain doing nonsense, which is exactly the right altitude for this stuff.