3rd Generation AI inspired by insects

Thomas Nowotny and Andy Philippides

Sussex University, Department of Informatics

Insects are expert at many tasks that are currently difficult for robots. In contrast to AI methods such as deep learning which can take a long time to train and need large amounts of labelled data, insects learn rapidly and robustly with tiny brains. For instance, the honeybee, a well-studied animal with a brain of only 1 million neurons, exhibits sophisticated learning and navigation abilities utilising highly efficient neural processes. We posit that these remarkable capabilities come about because learning is an active process scaffolded by innate behaviours which have evolved alongside eyes and neural structures enabling complex behaviour to emerge without complex processing.

At Sussex we are following this embodied approach to try and understand insect behaviour and find biomimetic solutions that can be implemented on robots by mimicking the sensors, neural circuits and behaviours of insects. In particular, we implement our models as spiking neural networks that are simulated using our GPU accelerated spiking neural network software which can be deployed equally on HPC and low power edge computing devices including on board small robots. In this talk, we will discuss how we have used this approach to develop efficient visual navigation algorithms that allow robots to robustly navigate outdoor routes with all navigational information encoded in a small neural network.