Loading…
Press J K L as a note reaches the line. Your opponent is a Bio-RL fly that has already practised this song 30 times, playing with its learning frozen.
Hit rate after each song
Each fly practised for 30 songs (the same song 30 times), starting untrained. Results are averaged over 5 repeats; song 0 is before practice.
What the options mean
How the fly learns
- Bio-RL
- The brain adjusts its own connections. Dopamine strengthens them after a hit and weakens them after a mistake.
- Deep-RL
- The brain stays fixed. A separate AI program learns which key to press from the brain's output.
How the brain is wired
- Real wiring
- Copied from a map of a real fruit fly's brain.
- Shuffled wiring
- The brain's overall structure (eyes to memory to output) is kept, but individual synapses are rewired at random. Tests whether exact connections matter.
- Random network
- Any neuron can connect to any other, so signals don't follow a real brain's route from the eyes, through memory, to output. Tests whether brain structure matters.
- Dopamine blocked
- Bio-RL with dopamine switched off. Tests whether dopamine drives the learning.
Deep-RL, real wiring, 150 songs
Deep-RL given 150 songs of practice instead of 30, to see if more practice helps. It reached 98.5%, about the same as after 30 songs.
What the numbers mean
- Before practice
- Notes hit before any practice.
- End of practice
- Average of the last 5 practice songs.
- New song
- Notes hit on a song it never practised, with learning off.
In short: the brain-style learner needed dopamine and the fly's real wiring. The trained readout learned on every wiring, even a random one.
Each grey dot is the cell body of one of 139,662 neurons in a fly's brain and nerve cord (Janelia and Google male CNS), seen from above with the head at the top. The empty band below the brain is the neck: only nerve fibres pass through it, so it has no cell bodies to draw. When a note falls into a lane, that lane's visual neurons fire, then Kenyon cells, then the output neurons that choose a lane; each flashes as it spikes. Nerve cord neurons flash when the brain's descending neurons send signals down the neck. For the Bio-RL fly, a hit releases reward dopamine and a mistake releases punishment dopamine, which strengthen or weaken the synapses that just fired.