Fruit fly brain model for making efficient future AI systems

Nervous systems of insects inspire machine learning and AI applications


Study explores functions of fruit fly's nervous system in food seeking / results valuable for the development and control of artificial intelligence.

Transformation of sensory information into memories in the brain could inspire development of future machine learning and artificial intelligence applications for solving complex tasks.

Zoologists at the University of Cologne analysed how insects learn to associate sensory information in their environment with a food reward, and how they can recall this information later in order to solve complex tasks such as the search for food.

The team studied the nervous systems of insects to investigate principles of biological brain computation and possible implications for machine learning and artificial intelligence. The study has been published in the journal PNAS.

Living organisms show remarkable abilities in coping with problems posed by complex and dynamic environments. They are able to generalise their experiences in order to rapidly adapt their behaviour when the environment changes. The zoologists investigated how the nervous system of the fruit fly controls its behaviour when searching for food.

The theoretical principles underlying this model can also be used for artificial intelligence and autonomous systems. They enable an artificial agent to learn much more efficiently and to apply what it has learned in a changing environment.

Using a computer model, the researchers simulated and analysed the computations in the fruit fly's nervous system in response to scents emanated from the food source.

'We initially trained our model of the fly brain in exactly the same way as insects are trained in experiments. We presented a specific scent in the simulation together with a reward and a second scent without a reward. The model rapidly learns a robust representation of the rewarded scent after just a few scent presentations and is then able to find the source of this scent in a spatially complex and temporally dynamic environment,' said computer scientist Dr Hannes Rapp, who created the model as part of his doctoral thesis at the UoC's Institute of Zoology.

The model created is thus capable to generalise from its memory and to apply what it has learned previously in a completely new and complex odour molecule landscape, while learning required only a very small database of training samples.

“For our model, we exploit the special properties of biological information processing in nervous systems,” explained Prof Martin Nawrot, senior author of the study.