Mathematically minded bees inspire neural network training
Scientists have demonstrated that replicating the neural networks of honeybees – which have the cognitive ability to grasp basic mathematical logic – could reduce the effort involved in training AIs.
According to the researchers, based at the University of Cologne, honeybees can perform numerosity estimation, which allows them to solve some simple mathematical problems.
“Experiments showed that insects such as honeybees can actually 'count' up to a certain number of objects. For example, bees were able to compare sets of objects and evaluate whether they were the same size or whether one set was larger than the other,” said doctoral student Hannes Rapp. For example, a bee was capable of recognising that six diamonds are “more than” four circles.
Rapp and his supervisor, Professor Martin Paul Nawrot, demonstrated these abilities in a computational model inspired by the honeybee.
So far, it has not been understood how the neural network for this cognitive ability is constructed in bees’ brains. Scientists had previously proposed a firmly implemented circular circuit with four neurons corresponding to the four operations “equal to”, “zero”, “more than”, and “less than”. However, Nawrot and Rapp found that just one neuron is sufficient.
The action potential (burst of electrical activity in the brain) of the single neuron varies depending on the mathematical problem the bee is faced with. This can be trained on the neuron. The researchers used this to build a computational model in which an artificial neural network (ANN) can be trained to solve mathematical tasks in a similar fashion. They found that when mimicking a basic mathematical task similar to that previously given to bees, the model achieves a similar success rate as the bees.
According to Nawrot, this approach helps the ANN learn: “A lot of money has already been invested into training [ANNs] to visually recognise the number of objects. Deep-learning methods in particular enable counting by the explicit or implicit recognition of several relevant objects within a static scene,” he said.
“However, these model classes are expensive because they usually have to be trained on a very large number of patterns in the millions and often require cloud computing clusters. Our honeybee-inspired approach with a simple model and learning algorithm reduces this effort many times over.”
The researchers recounted their work in an iScience paper.
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.