Animals such as fruit flies or zebrafish have so far been regarded as comparatively simple stimulus-reaction machines that react to certain environmental stimuli with fixed behaviors. However, scientists at the Max Planck Institute for Biological Cybernetics in Tübingen have found evidence that there are amazingly long-lived internal states in the brains of these animals that control behavior. This could not only explain the amazing flexibility of biological brains, but also help AI research further, reports Technology Review in its December issue (can be ordered at the kiosk or online).

Biological brains are networks with strong feedback, which allow very complex stimulus-response schemes to be developed. However, the number of possible schemes is also astronomical. Drew Robson and Jennifer Li from the Max Planck Institute for Biological Cybernetics in Tübingen suspect that internal conditions direct behavior into certain corridors – and thus could be a way to drastically reduce the complexity of the problem.

Changing states of activity Robson and Li examine recordings of the brain activity of fish larvae. With the help of fluorescent neurons, they were able to show that the patterns of brain activity of the larvae differ systematically during exploration and hunting. They were also able to identify a group of nerve cells in the center of the fish’s brain that was active as long as the animals were in hunting mode – even when there was no prey nearby. The larvae only stopped hunting when the cell group was no longer active and switched back to exploration mode, which was accompanied by its own sustained activity pattern in the brain.

AI pioneer Jürgen Schmidhuber, scientific director of the Institute for Artificial Intelligence in Lugano sees parallels in this with the function of certain neural networks. “There have long been artificial neural networks that can switch between different states of activity,” he says. “More recent examples are the artificial video players for Dota from OpenAI and for Starcraft from DeepMind.” In millions of games, these algorithms learned to control various agents in strategy video games – sometimes on a human level.

“The progress made by AI in strategy games like Starcraft is remarkable,” says Robson. “But it’s also remarkable that human players are competitive without having played millions of games.” Evolution has optimized biological neural networks for survival over millions of years. “It wasn’t just the number of neurons that grew, but also their careful coordination and specialization,” says Robson. Internal conditions could play a central role in this fine-tuning.

(bsc)