How a brain the size of a sesame seed could change AI forever
- 28 September, 2018 08:08
When a honey bee leaves its hive for the first time in its life, it will typically perform one to five flights around the surrounding area to get its bearings, then get straight to work.
“Their learning phase is negligible, or close to negligible,” explains Macquarie University’s School of Biology director and bee-obsessive Dr Andrew Barron.
Compare this with a drone, armed with computer vision and deep learning. Last year, researchers ‘taught’ one to fly around the hallways of their lab facility. Just learning how to not crash into things took the UAV 40 hours of flying time and 11,500 collisions.
It’s taken more than collision-free flight and the ability to return home for the bee to survive and prosper for the last 130 million years or more. Despite having a brain only the size of a sesame seed, they are capable of some impressive cognitive feats.
“Even if we ignore all of their social behaviour, they have concept learning; they have lifelong memory; they can plan; they can plot incredibly efficient foraging routes over kilometres, finding these ephemeral dispersed resources, and get them back to the colony – and make a profit. And their navigation is superlative,” Barron gushes.
Researchers at the Queensland Brain Institute have even suggested bees have a level of self-awareness, “if not consciousness”.
“In terms of robust cognition honeybees are astonishing,” Barron adds.
Applying a deep learning model to just one of the tasks that comes naturally to a honey bee, say, reliably identifying the pollen heavy flowers in a meadow, would take millions or even billions of training examples, and a significant amount of compute.
Bees do all they do with less than a million neurons (by comparison a mouse has around 75 million and a human some 100 billion). For Barron and others like him, this presents a tantalising possibility that could give rise to a completely new form and philosophy of artificial intelligence.
“I’m building a computer model of the honey bee brain,” he says
And his ambition doesn’t stop there. “That’s the way we build towards modelling the human brain,” Barron adds. “We start simple and we build up.”
Sting all humans
Insect-brain-inspired AI is a slightly fringe field, Barron admits. But it presents potential advantages over deep learning in many applications.
“I’m not in any way dissing deep learning. The progress it’s given us is astronomical and very impressive. But for me as a neuroscientist there’s a very interesting point of comparison: the kind of brains I study do not use deep learning in any way at all, but they achieve robust, flexible, efficient cognition,” he says.
One of the biggest advantages from tying computer models to biological brains is that insect and animal intelligence has built-in limitations.
The actions of many artificial intelligence systems are difficult to predict and manage. When Google’s AlphaGo AI beat the world’s best human player, for example, it made a number of what the commentator described as “not a human” moves. It is susceptible to reinforcing its own biases, often the result of its input data being historic, rather than observed in real-time.
As Barron puts it: “No honey bee has ever gone Skynet and decided they would kill all humans”.
“We want to build AI which has limits as to what it can do; limits which are intrinsically part of the architecture. In terms of deploying things that are up in the air or move freely through the environment, having forms of AI that have intrinsic limits is not a bad thing,” he says.
“If we want something that could operate autonomously and with cognitive flexibility and yet be safe and trustworthy, that’s when a bee brain model could be a very useful model to apply,” Barron adds.
Explaining the inner workings of the neural networks behind deep learning systems is a challenge that continues to confound researchers. The likes of IBM, Microsoft, Accenture and Google have released products to help businesses shine a light in the so called 'black box', but these generally only analyse inputs and outputs.
Since bee brain models don’t use deep learning it might be easier to run a post-mortem on, say, why a drone controlled by one had crashed, Barron says.
“In theory it would be easier to diagnose or perform an autopsy on a problem in a drone with this kind of control system then one with current deep learning control systems,” he says. “We can interrogate them, so we should be able to understand why and how the failure happened.”
Autonomous, self-organising, decentralised
Reverse engineering insect brains sounds wildly ambitious. But Barron and his colleagues in the field – who are currently seeking to establish an Australian Research Council Centre of Excellence for NeuroRobotics – are not starting from scratch.
“Some of the earliest microscope drawings ever made were of the honeybee brain,” Barron says. “We have 200 years of neuro-anatomy behind us. We don’t have a connectome of the bee brain but we have a really good understanding of what connects to what. We already know enough that we can start to make simple circuit representations of key regions.”
The honey bee brain is also “incredibly modular” with “very specific bits that do very specific jobs”, he adds.
In a paper published this month, researchers from Macquarie University and the University of Sheffield in the UK demonstrate that a computer model they created, inspired by the part of the honey bee brain responsible for abstract concept learning, was able to learn concepts such as ‘sameness’ and ‘difference’.
“It’s the first concrete, neurobiological model of abstract concept learning that doesn’t imagine ‘a blob that does it somehow’,” Barron says.
In another project – called Brains on Board – researchers are reproducing neural models of bee navigation and action selection, and using them to fly drones.
The potential is huge, Barron says, particularly for agriculture and mining.
“Cases where we need to gather dispersed, hidden, hard to access resources, and bring it back to a central place, in environments that are not pleasant or hazardous for humans – this is exactly the class of problems that ant and honeybees and social insect colonies have solved so efficiently that they’re the dominating biomass on the planet,” Barron says.
“And they’re doing it in a completely autonomous, self-organising, decentralised way. We can learn from that, and safely deploy it in ways that we are not currently imagining.”