Q&A: How bee brains could inspire AI in flying robots
Researchers looking to improve artificial intelligence modelling via bee brains
Researchers at two UK universities are trying to use the guidance systems and intelligence inside the brain of a bee to improve artificial intelligence in autonomous flying robots.
According to the researchers in Sheffield and Sussex, the ultimate goal of the project is to be able to send drones or micro UAVs into action in dangerous locations, such as searching for survivors in collapsed buildings. But AI is still too slow and complex to react quickly to the environment in front of it and alter controls accordingly.
The AI for autonomous controls has long troubled automation experts, but the scientists believe modelling their systems on brain power found in nature could solve many issues – especially given the rise of parallel computing in graphics processors, which can be repurposed to model brain patterns.
If we could make a flying robot as intelligent as a bee, we might be able to give it tasks that humans would find difficult or dangerous
The idea of mimicking nature in AI research is nothing new, but according to James Marshall, lead researcher on the “Green Brain” project, the bee's brain is a more manageable blueprint for building a model of a brain and could provide a real breakthrough. We got in touch to find out why.
Q. How could the research help AI development in the long term?
A. I would say that AI has a long history of over-promising and under-delivering... we were supposed to have computers as clever as us by now, but this hasn't really happened. The early approaches were based on things like logic programming, trying to engineer something as intelligent as a human, but increasingly people are realising that by studying how animals behave, and the mechanisms that give rise to these behaviours, we might be able to do better.
We were attracted to the honeybee because, as honeybee experts and neuroscientists are showing, it's actually very sophisticated and able to do a lot of things that animals with much larger brains do. But, given it has only a million neurons, building good models of the honeybee brain and simulating them seems much more feasible than modelling, for example, the 100 million neurons in the rat brain.
Q. Does the lower brainpower impact the usefulness of simulated systems?
A. Researchers who study insect behaviour and neuroscience argue that the difference between brains like those of honeybees, and vertebrate brains such as those of rats, is one of quantity rather than kind. Their argument is that a lot of the cognitive sophistication is already observed in the insects; the need for a bigger brain comes, at least in part, from the need to control a larger body. We hope that because of this, what we learn and what we're able to develop will actually be quite general and powerful.
Q. Is the ultimate aim to build an autonomous robotic flying device?
A. The aim is to understand bee cognition better, and to test our understanding by embodying our brain model in a flying robot. But, in the longer term, autonomous flying vehicles have a lot of obvious applications... If we could make a flying robot as intelligent as a bee, we might be able to give it tasks that humans would find difficult or dangerous, such as searching for survivors in collapsed buildings, for example.
Q. To what extent would such a project rely on hardware developments, or is the AI component the missing link?
A. The hardware is really just off-the-shelf. We plan to make quite minimal adaptations to the kind of flying quad-copters you can buy for a few hundred dollars, and then fly around from your iPad or your phone. These have onboard cameras and Wi-Fi transmitters/receivers, so they are perfect for our purposes. All the really clever stuff will take place in the bee brain model, running on the very fast GPU-accelerator cards that have come out of the 3D games industry, and are now becoming increasingly important in scientific computing.
Q. How does parallel computing help in this type of research?
A. In order to control our flying robot, our brain models need to be able to run in real time, receiving a video stream and outputting appropriate motor commands. Split seconds of delay could make the difference between successfully achieving a goal, and crashing into the floor.
The only way we can run our brain models in real-time is by using parallel computation; fortunately a real brain does a lot of parallel computation, of course, with areas that work largely independently with few connections between them, but many connections within. So, implementing the brain simulation on a parallel computer becomes feasible, but still an interesting and substantial technical challenge.