Since we know of no better thinking machine than the human brain, one of the main objectives of machine learning is to build an artificial copy of it. But while some very advanced machine learning algorithms have been developed in recent years, none of them are actually very much like the brain. By comparison, they are very slow on the uptake, easily fooled, and terribly inefficient. So why not just clone a brain and declare that artificial general intelligence has been achieved already? Surely we have more than enough GPUs to simulate all of the neurons.
That is much easier said than done. The problem is that we do not understand how the brain works well enough yet. A team led by researchers at the University of Sheffield wanted to fill in this gap in knowledge, but given that the human brain is extremely complex, they decided to start a bit smaller. They created a computational model of the sesame seed-sized brain of a bee. By using this model to better understand the function of bee brains, we can glean some insights that will help us to improve our algorithms today, and perhaps ultimately get us to a better model of the human brain.
The neural network model of a bee’s active vision (📷: H. MaBouDi et al.)
In the course of their work, the team found that bees do not just passively observe their environment. Rather, they actively shape what they see by moving their heads, bodies, and eyes in strategic ways. These flight movements create distinctive electrical patterns in their tiny brains, making it easier to extract meaningful information from the visual chaos of the natural world. And somehow, this tiny system can solve difficult visual discrimination tasks, such as recognizing human faces, with far fewer neurons than any artificial system in existence today.
The researchers used this insight to construct a highly efficient, biologically inspired digital brain. They then tested it with a range of challenges, including a pattern recognition task where the model had to distinguish a plus sign from a multiplication sign. Just like real bees, the model improved its accuracy dramatically when it mimicked natural bee scanning behavior.
This suggests that movement is more than just about getting around — it is also an integral part of how animals learn. Rather than brute-force number crunching, intelligent systems might benefit more from smart sampling: moving to see better, to think better. The bee model’s neurons gradually adapted to the motion patterns of the visual input, forming efficient, sparse codes that required minimal energy. Unlike standard AI models, this one used non-associative learning in which it refined itself without needing constant reinforcement.
The model mimics how bees scan a scene visually (📷: H. MaBouDi et al.)
Furthermore, the researchers found that active scanning helps encode information in a compressed and efficient form in the bee’s lobula, a visual processing center. When paired with additional neural structures that mirror the mushroom body (which is used for associative learning), the system performed well across a wide range of visual tasks.
Ultimately, this study might offer us a roadmap to smarter, leaner AI. If we want machines to learn with the efficiency and elegance of natural brains, we may need to start thinking not just about what they see, but how they move.