Transcript of journalist and senior media executive Richard Sergay's interview with Dr. Andrew Barron and Dr. James Marshall for the "Stories of Impact" series.

Watch the video version of this conversation.

 

RS =  Richard Sergay (interviewer)

AB = Andrew Barron (interviewee)

JM = James Marshall (interviewee)

 

AB: I'm Andrew Baron, Associate Professor at Macquarie University in Sydney. And a Leverhulme visiting professor at Sheffield University.

 

RS: Diverse intelligence?

AB: Define diverse intelligence. Really interesting question, which is probably close to the motivations of the project. We, intelligence is one of these things where we have a clear intuitive sense that it really, really matters. Putting the envelope around what intelligence is, is extremely difficult and I think what will help us frame that envelope is if we can accept the diversity and study the diversity. So if we study intelligence not just in humans but in other living things, potentially even other machines, we can tidy up that definition of what intelligence is, what differences in intelligence make a difference, and where we draw the boundary on what is intelligent and what's not. 

 

RS: How difficult is it to get at what intelligence is.

AB: Getting at what intelligence is I think is conceptually extremely difficult but I think there's also points of tension there that we just have to be comfortable with. There's a very famous and I think very, very informative analogy which is how could we compare the amount of flight in something. So we have a jumbo jet. It's great at flying. We have a bumblebee. It's great at flying. We have an osprey, great at flying. They all fly in completely different ways and I could not say which one has more flight-iness or is a better flyer than the other because they're different. We're okay with that tension because we've got fundamental principles of wing loading and aerodynamics that let us talk about the kind of different flight in these different things you know in a meaningful way that lets us understand the differences. In terms of defining intelligence I think we need to move towards that. If we have a better sense of what matters for different intelligent capacities then we will be able to talk about intelligence in a way that we can be comfortable without properly defining tightly what intelligence is. Because we'll have a sense of the shape of the bounds of what are the properties of intelligence.

 

RS: Tell me about your project, bee brains in particular?

AB: So my project is completely awesome, I love it. My project is particularly focusing on honeybee intelligence because it gives us such an informative lens, sort of informative comparative lens on the intelligence of other animals including humans. We have these tiny little animals with really minute brains that to us are minuscule that we can see them, but they're milli-metric scale. They have a million neurons. It's minute compared to a human brain. But you only have to spend a bit of time in the presence of a bee to understand that you're in the presence of a very cognitive agent that is actively deciding and choosing moment by moment how to organize what is a, intellectually very demanding lifestyle of having to gather pollen and nectar from all these dispersed flowers and not run out of fuel and make a profit and get it back home. I'm ranting, I always rant when I talk about bees but that's because how they do this, how they achieve this is what really fascinates me. And because their brains are so small. This question of how they do these kinds of things. It's a tractable question, it's something that we can actually get to grips with and grasp and take apart and try to understand. And then as we understand that all these really interesting intelligent things like complex learning, complex memory, complex navigation, complex assessment, we'll learn some evolved solutions for that and we can then ask, is the human brain doing this in a similar way. How have we evolved similar solutions to these kinds of problems? And it's going to give us what we urgently need, which is a lens into how our intelligence works and how our brain works. 

 

RS: The genesis of this project?

AB: Ooh, the genesis of the project is-- how far back do you want me to go?

 

RS: Your interest in insects and bees in particular?

AB: Okay so that's, the interest in bees was very specific. So my PhD which was last century, end of last century, a long time ago, was on the Drosophila, the fruit fly, and I was studying learning and memory and fruit flies. It was a good PhD but what I learned was that their behavior wasn't at the level of complexity that interested me. Fruit flies could do certain forms of learning but not the kind of exciting cognitive stuff that relates to the kinds of things that we do, like concepts and, and complex spatial calculations and complex assessments of relationships. Honeybees could do all of that. And their brain is only 10 times bigger, it's still in this scale where it's analyzable. So for me the bee was in this unique position where its behavior was complex enough to be interesting but its neurobiology and its brain was simple enough that we could study it and it had a complete genome sequence. And when I started this project and started this interest there was nothing else in that space. So for me the bees gave a unique opportunity for what interests me, which is the mechanisms of behavior. It's not what animals do, it's how they're able to do them with the new biological resources that they have. 

 

RS: You mentioned the genome. How important was it in terms of the work you do?

AB: What the genome information gave us is what I needed to really justify the comparative dimensions of my project. The genome counter-intuitively told us that the honeybee is closer genetically to a human than it is to a fruit fly. Which sounds weird but that's just the way evolution works sometimes. So all of the proteins that matter that make a difference to the operation of neurons and the way that define how neurons function, the similarities between the neurons in a bee and a human massively outweigh the differences. The differences become detail. The similarities tell us that they have the same operational principles. So what this means to me is that in terms of the bee brain and the human brain, yes, they're very, very, different. But the substrates are similar enough that the insights that we learned from the bee will translate to the vertebrate lineage and that matters to me.

 

RS: Why the honeybee in particular?

AB: Of all the bees, the honeybee is the easiest to work with and has the greatest background of study. So we know the most about honeybee behavior, honeybee natural behavior, honeybee neurobiology, and honeybee brain anatomy. We have decades of wonderfully detailed study of the honeybee. So for a project like mine, I'm not starting from scratch. I've got this incredible depth-- breadth and depth of knowledge that I can draw on for this project. 

 

RS: What is unique about our honeybee and what it does?

AB: What's unique about the honeybee? So-- many things are impressive about the honeybee. But what stands out as a unique feature of the honeybee would have to be its symbolic dance language. I have to be careful here because there are other examples of complex-- many other examples of complex communication in animals. But the dance language is compelling and there's nothing else quite like it. And what sets the dance language apart is that they are transforming information about distance and direction to things in the real world, to these remote food sources, into a single vector that they can signal through a dance. So this spatial information becomes an element of a highly stereotyped dance movement. And what is astonishing is that other bees can backwards interpret that dance and translate a movement in the dance to a flight vector in space. That's, all honeybees do this but no other bee does that, that really does stand out to me as a remarkable example of an evolved behavioral innovation. 

 

RS: Dogs or cats versus a honeybee? Why honeybee?

AB: There's a lot of answers to that so I might get lost in my own answers. But to try to address the question of why a bee, why not a mammal. There's a lot of benefits to working on a system like a bee. So one of them is the scale of the system. So in a bee brain, I have one million neurons to work within my system. A large part of my project and the reason I'm here in Sheffield is that we're trying to create a computational model of this bee brain. Now doing that with a million neurons. And given how much we know about what the different parts of the bee brains do, we can very quickly constrain our modeling effort. We can on a whiteboard draw up the circuits that we think are working in a bee brain and with just a couple of networked desktop computers we can create a computational model of this theoretical model. I can't do that with a mammalian brain. So even the smallest mouse brain, we're pushing up well into 80 million neurons. You get towards dogs and cats and it's far, far, far bigger still. And the neurobiology is much less constrained. We know, in fact there's still enormous debates about specifically what the different regions in the -- in the mammalian brain do. That makes modeling the whole brain very, very difficult. So where we are with the state of the art is we're modeling bits of mammalian brain. But with insects, we're ready now that we can model the whole thing and that's what the work I'm doing in Sheffield is about. 

 

RS: Why is a computational model important?

AB: A computational model is, it's building a circuit diagram of the brain in a virtual world. So we can draw up how we think neurons connect and the kinds of connections that they make. With a computational model we can implement that and we can then make it a dynamic system that we can feed input to. It will process the input in the way that we think that the honeybee brain is processing it and it will give us an output. And we can analyze that output in terms of what is the system doing, what the bee's doing. If it is maybe our model is close to reality. If it's wrong then we've learned something, we've learned that our assumptions were wrong and we have to recast and start again. It's the dynamic element of a computational model that matters because often when you run these systems and feed them real world input, they do behaviors that you wouldn't predict. But behaviors that then suddenly really closely match what the animal would actually do.

 

RS: You call them spectacular learners. Why is that? 

AB: The honeybees really are spectacular learners. They-- they learn very fast and very robustly. As an example, if we give a honeybee something simple to learn like this odor is associated with nectar. This odor is where you find nectar. It will learn that on one trial if you give it three trials it will learn that for the rest of its lifetime. So that's very fast acquisition of relationships between information. But they're also able to map these associations across really quite complex relationships. Which means that they can even learn things that we would consider to be abstract concepts, things that we would call learning of sameness, learning of difference. A honeybee is able to do that. That hasn't been shown in any other invertebrate that I know of. 

 

RS: What kind of abstract learning in particular?

AB: Abstract relationships that we've been studying particularly are the relationships of same and different. So you're not learning any particular stimulus, you're just learning I need to match something or I need to not match something. This concept of matching is one that we're studying here in Sheffield. 

 

RS: Example?

AB: An example would be I give you a flash card of the ace of spades. I hold up two cards. One's the ace of spades, one's the king of hearts. And all you have to do is tell me what's the same so you're not learning any card in particular. I could show you any cards that you can apply this rule of the same. And you could solve the task. 

 

RS: And for the honeybees?

AB: For the honeybees we had a binary choice--

 

So for the honeybees we had a binary choice where we showed them a wave shaped maze with a blue color and a yellow color for example in the two different arms of the maze. But on entering the maze we showed them either blue or yellow. And if yellow was at the start they have to go to yellow, if blue is the start, they have to go to blue. And we can keep changing that out with any other stimuli and they can still solve it.

 

RS: Cognitive characteristics that you're seeing in honeybees that you thought were restricted to mammals?

AB: The list is ever growing. So there's a whole list of cognitive-- cognitive facilities that we thought these set humans and mammals apart from simple animals. Then we look at the simple animals and we find that's not true. We're finding these things here. This, the study I talked about, the study of abstract concept learning, that was one of the real pivotal moments when we thought there's an awful lot more that this insect brain can do than we realized. Things that I've studied directly is metacognition in bees. Which is honeybees can actually adjust their behavior in response to their state of knowledge about a choice. Now again this is a high order cognitive concept that was again thought to be limited to first of all humans and then humans and other higher mammals. And yet even honeybees can solve this. But again what really motivates me in this work, it's not creating the list of what the bees can do. It's using the simplicity of the bee brain to ask the more insightful question of how is the bee doing this. If the bee is solving a task that we think demonstrates metacognitive How can an animal with just 1 million neurons do that. To me that's the much more-- greater insight, because what it force-- it forces us to rethink our assumptions about what we would need for medical mission. It forces us to take a computational perspective on the problem and ask what is the minimal computational architecture that could do this and then you realize a lot of these things that sound complex in terms of computation are not as complex as we might think. And that's what then gives us this real genuinely beneficial insight into diverse intelligences, because sometimes our assumptions about what is intelligent are misled.

 

RS: Define metacognition?

AB: So meta-- an example of metacognition is an awareness of the state of your own state of knowledge. So the statement I don't know is an example of matter cognition. You're assessing a circumstance and you're coming to the conclusion that you don't have enough information to address that or to answer that. So then rather than guessing you will seek more information.

 

RS: And its application to honeybees?

AB: We set bees exactly that task. We trained them again in a choice but we made some of the choices impossible to solve and we made some of the choices possible to solve but really, really subtle distinctions. And we also gave them an option where out of any one choice they could opt out, so they weren't forced to choose. They could decide to leave that trial and take the test again with new stimuli. So we gave them the option to express I don't know by not choosing and by exiting and deciding to try again. 

 

RS: And the experiment, describe it to me. 

AB: We had the bees flying through a choice chamber. And in the choice chamber we presented them with these two stimuli and the stimuli could be completely different. There could be no difference or there could be very perceptually subtle differences between them. In between the two stimuli is what we call the opt out port-- port. So the bee could make a choice and if it made the choice and got it right it got sugar. If it got it wrong it got quinine which is not harmful but it's bitter tasting. Or they could just fly straight on through and not make a choice and opt out. If they flew through they could come back to the start. In the meantime we reset the test with new stimuli and they could try again. And when we reset with new stimuli they might get an easy test this time rather than a hard one. So we gave the bees a chance to express I don't know. And to opt out rather than forcing them to choose. 

 

RS: And what did you find?

AB: We found, and this was actually an awful lot of work but it was it surprised us. We found that when we gave them a difficult test or an impossible test they opted out far more than when we gave them an easy discrimination. So they were using the opt-out port far more when the discriminations were hard to do. Now that alone isn't enough to say the bee is showing meta cognitive processing. Because they could just be learning that, they could just be basically biasing the use of the opt-out port to any discrimination that is difficult. The clincher and this is this takes a while to get your head around. The clincher was that we compared how the bees did in the difficult tests when we gave them the chance to opt-out or we gave them no chance to opt out. We forced them to choose in difficult trials. When we forced them to choose in difficult trials, they basically performed at chance level, just slightly above chance level. It really was hard for them and they were only getting it right just a fraction more than they were getting it wrong. When we gave them the chance to opt out, if they chose to take the test they actually performed much better. So they weren't just opting out at random when the test was hard, they were opting out selectively when they assessed they had insufficient knowledge to solve that particular trial. And that meets an operational definition of metacognition. Now what's been equally exciting is working out how we could get the honeybee as a neural system to actually perform that task and it's changing the way we think about meta cognitive processing. 

 

RS: Bees exceed rats in performance tests in learning and cognition?

AB: So that's yes, that's partly my, I've got to give the bees a wallop on the rats since in many ways. It is actually, if we look comparatively across the literature, in many tests, even these tests of very simple learning or even tests of very complex learning we see the bees learning faster than rats. I don't have an answer for you as to why that is yet. It fascinates me. We have an organism where our assumption is that this is smarter. And yet in a whole battery of tests of learning, tests memory, tests of spatial cognition, the bees are outperforming the rats. So most, the most recent example and this is work I haven't published yet, but it's work that Templeton funding and allowed us to do, we set the bees what was called a contextual discrimination task. Which is actually really quite hard. We had two stimuli one rewarded and one punished, but we also had two contexts which the bees saw before they entered the test. So in contest-- context one, let's say this was awarded this one's punished.

 

In training we constantly flip between the two contexts. Context 1, context 2, we flipped it randomly. Whenever the context changed, which stimulus was rewarded changed. So this stimulus is rewarded in context 1, this one in context 2. And through the training which is constantly randomly changing that relationship. So the bee can never learn that one stimulus is rewarded, it's equally rewarded as punished. It's got to use the context to learn the rule to identify when this is rewarded and when this is punished. The bee solved this far faster than I thought that they would pick it, they solved it in about 10 to 15 trials. When we look at comparative rat data, they're taking 50 or 60 trials to learn a similar task. There's something about that difference that matters and it matters in a counterintuitive way. But when we understand it, then again we'll be learning about these dimensions of intelligence that matter. In this case it may be that the simplicity of the bee brain is actually helping them learn these elemental relationships. 

 

RS: Did you shoot any video?

AB: We have endless videos of them.

 

RS: Did you send us any of that?

AB: Not yet. But I can. 

The quality won't match-- and it's nice because we've got bees in 3 dimensions coming in and making complex choices. We've got loads. 



RS: Where are we in understanding the bee brain?

AB: I suspect I'm not alone in saying this but I think that in the arc of understanding that the bee brain, we're at the most exciting point. We're at a point where we're really getting a synthesis. So in, up until quite recently we've had really good information about some bits of the bee brain, not quite so much information about other bits of the bee brain. We're now at this point thanks to advances in connectomics and thanks to continually improved methods in neural reconstruction and microscopy and computational methods, we're really getting to the point where we can put not just the bee brain but insect brains together as an information flow system. And very, very soon we'll have very detailed connection terms for a number of insects and then we will have very, very detailed neural biological reconstructions of insect brains. This is important for a number of reasons but particularly for the project time involved in modeling the bee brain. We need that information because the information about the connections in the bee brain, that helps us constrain our models. So our models are bound to biology. And it restricts the space of our models in a realistic way. So we really are at the point now where we can meaningfully create system-wide models of the insect brain and this is going to, it's emerging now, it will explode in the next five years.

 

RS: Connectome?

AB: A connectome basically means a detailed map of the connections between neurons in a bee brain. There's various resolutions you can do that. So the term again is various interpretations but it basically means it's this physical map of connections in a brain. 

 

RS: And that helps us understand what exactly?

AB: The physical map of connections in the brain, that constrains the way that information can flow in the brain. It doesn't tell us anything by any means but it's certainly extremely important information for modeling structures, modeling systems in a brain. 

 

RS: Modeling-- the implications?

AB: If we get that right, to me there's two implications. For me personally the most important implication is we have a tool. So we have a tool that can let us really critically explore hypotheses about how brains might actually work. Because in the model we can do the kind of manipulations on the model that we would love to do in the brain. We could change any element and see what the consequences of a brain function. We could take a honeybee brain we could triple the size of it and ask how does that change how it works. We could do that with a model of the brain. So it gives us a fantastic way to look at fundamental questions of how brains function. It's also a bridge. So if we can model the bee brain, we can take insights from those models and translate them directly into technological applications. If we can model the bee brain all of this intelligence, all this dynamic autonomous behavior that we get out of bees, we should be able to capture that in the model. Now we wouldn't just then say ok, we've got a system that will work for robotics, but there'll be things that we can learn from that that we could translate into robotics. 

 

RS: Such as?

AB: Bees have evolved for millions of years to be fantastic autonomous behavioral control systems. They're really robust, they're really reliable, they solve very, very complex image processing systems. They're amazing navigators across very large distances. All of these are current challenges in autonomous robotics and yet the bee is doing it with incredible computational efficiency. So if we can solve how the bee is doing it, there will be insights there that we can learn that we could translate across to make robotics more efficient, more robust, also more comprehensible. Systems that if we start with a deep understanding of a system like an insect brain, we haven't understood system that we can apply in robotics.

 

RS: Military applications? 

AB: It's inevitable that the military would be looking at something like this. That's not something that interests me. What interests me is the capacity for safe robotics. I think what all of us want, whether we like it or not we're in this robotic revolution. It's happening. It will only accelerate even further. What we want from robotics, we want devices that are useful and that help us. But we need them to be trustworthy and we need them to be safe. If we're going to have a system that is trustworthy, we need to understand how that system works very, very, very deeply, so that we can understand how it works, we can also anticipate how it might fail and the kind of failure it might make. If we're starting our robotics systems on the basis of a deeply understood system like the bee brain, which is, has a form of intelligence with the intelligence is limited by the structure of its brain. If we could capture that and move that into robotics, to me we have a system that is more intrinsically understood. And I think therefore potentially safer and more trustworthy than some of the current approaches in robotics. 

 

RS: Put yourself in bee brain. Describe what you are seeing and doing?

AB:  So you're kind, you're kind of asking me to ask the question, what is the experience of a bee. And this is something that interests me a great deal. It's a question that is philosophically hard to grasp. But I think that the only way that I can describe this and then you'll realize how it's philosophically hard to imagine what it's like. I think the level of experience of a bee is it's experiential of its world but unable to, unable to frame the question of what the world feels like for it. It doesn't have that self-consciousness but I think it is operationally aware of its surroundings and it's that awareness of itself in the world that allows the bee to do everything it does so well and so robustly. But I don't believe the bee brain's capable of expressing what the world means or what the world feels like for the entity. So it's very hard for us, because we're bound to this, we constantly ask questions about what would that feel like for me to be a bee. I think the bee's actually below that level of processing their experience of the world, but they're not experiential, they're not self conscious of the world. 

 

RS: You said that insects have already solved issues of robustness, stability, adaptability, in autonomous behavioral control systems. Explain that to me.

AB: Wherever you look at a behaving bee, the bee moment-by-moment is making very, very complex evaluations about how it's going to invest it's time for the maximal benefit of its colony. Is it going to forage today, is the weather good enough for it to forage? If it's going to forage what flowers is it going to go to. Is it going to get pollen or nectar for its colony? If it's in a patch of flowers, is this patch of flowers good enough to justify staying and trying to find more, minute droplets of nectar, all hidden in these individual flowers. Or is it time to give up and take what it's got back to the hive. Should it decide to try another patch of flowers that probably had nectar yesterday. Which is in a completely different part of its environment. Incredibly complex moment by moment decisions, all happening on board within this tiny brain with no director, it's resolving all of these moment by moment by itself, fully autonomous control of an extremely complex information-based problem. Bees have cracked that. The success of bees in their environment has proved over millions of years how well they do this. I mean I always tell my students, imagine setting yourself the goal of being a human making honey. You would find it impossible, absolutely impossible. We would have no-- we would never be able to efficiently gather micro liter droplets of nectar from flowers in a way that we could make an energetic profit. Bees do that. It's an astonishingly complex task and they do that. 

 

RS: And that is the bees mission. 

AB: That's the bees evolved lifestyle and their brain enables them to solve that task. But all of the complex evaluations within that task are somehow being resolved by this minute cognitive microprocessor of less than a million neurons. Which is why it's so fascinating to study because if we're thinking about let's imagine autonomous drones that we could use in exploration or agriculture or in mining, collecting resources, or even recycling from landfill, collecting scattered di-- sparse resources where you can't predict where they are, they're valuable that they're (INAUDIBLE) and unpredictably. And you've got to optimize finding those looking after yourself and getting them back to one point and making a profit. That would be the holy grail for so much robotics. Bees have solved that with this minute brain. 

 

RS: If we can understand what's going on in the bee brain there is a possibility of programming a drone to do that. 

AB: What I think is that if we can understand how a bee does that, there will be principles, there'll be tricks, shortcuts, in terms of these decision systems, the organization of a stable behavioral control system, and in particular in very efficient computing. And I think those tricks are the things that we could translate into robotics to help us improve performance in robotics. 

 

RS: How far away are we from that understanding?

AB: I think it's fair to say that we're already making good progress towards that outcome. I see it not as a career spanning or generational spanning goal. I see it as a five to ten year goal. What's in the way is we need a better understanding of the insect brain but that's coming, it's coming all the time, and we need more time in this project. We need more time to analyze what we learn-- what we're finding about the bee brain and to translate it into robotics. But the work is already happening. 

 

RS: Templeton funding is allowing you to do what?

AB: The Templeton funding is allowing me to look at these fundamental questions of comparative intelligence. To really get a grasp-- to really get a grip--

The Templeton funding is really letting me look at these fundamental questions of comparative intelligence, to really get a grip on how smart a bee is and what are the limits to bee's intelligence and why are the limits there. So let's stop thinking about what these can do but let us understand how they can do what they can do and then what stops them doing more complex things. And that's what Templeton's allowing me to do. 

 

RS: The architecture of a bee brain compared to a mammal. Describe that for me. 

AB: The architecture of a bee brain is fundamentally different to a mammalian brain. All mammalian brains, they have a common kind of plan to them, they look different, obviously look different. There's a common plan. There's a common plan to the insect brain too, but it's completely different to the mammalian brain. Regions that we talk of a lot in the mammalian brain like the cortex, just not there in the insect brain. And there's not even anything like it. So it's a fundamentally different information processing system. And for me that point of difference is one of the strengths of working with a honeybee. We learn a lot about fundamental principles when we study different systems and the bee is different enough that it's really insightful. 

 

RS: How do you draw potential insight into a mammalian brain or a human brain by understanding an insect's brain?

AB: To draw insight between the two, this is actually where computational modeling is really valuable. Because we could get hung up on the differences in the anatomy of the bee brain versus the human brain. But what we can do is we can look at the bee brain and we can say, what computation is enabled in this structure. We can do exactly the same with bits of the mammalian brain. What computation is enabled in this structure. And that means that we can actually compare these superficially very, very different looking systems. And then what we find is that bits anatomically look completely different, they're solving the same sorts of computation, and that gives us grounds to compare and also starts to tell a really, really interesting story about how, how brains evolved and how different forms of intelligence evolved.

 

RS: Flight?

AB: So you are right. I mean we move through our worlds in very, very different ways. The honeybee flies, we don't fly, we walk. But both us and the Bee, we're visual navigators. So we rely on our visual senses to control our movement through the world. The bee does it when it's winged, we do it when we're walking. For that to work we need to be able to subtract our own movements from things that are moving in the world around us so we can tell if we're moving through the world or if something in the world is moving past us. We solve that problem computationally, bees solve that problem computationally with exactly the same principles. So in terms of behavioral control, the differences between flight and us, it's less of a difference than you might think. It's also one of the reasons why I like to study ants as well as bees, because ants have to walk through a cluttered environment. And so in that way there's more similar to mammals, similar to, similar to rats certainly. 

 

RS: Language? 

AB: The honeybees, and I've talked about this a little bit about that, but the honeybees have this symbolic dance communication. Now that is often described as a dance language. I think linguists will get annoyed by describing the honeybee's advanced communication as a dance language. There are features in dance communication which are not features of human language. So dance communication does not have syntax, for example. Dance communication does not have the arbitrariness that human languages seem to be able to have but still function. So we don't have language in the sense that we have human language anything like a bee, but that's not to underestimate the sophistication and the elegance of dance communication as an information communication system.

 

RS: Other ways they communicate?

AB: Communication honeybees are very, very rich. Dance is pretty the most famous mode but there's also an entire array of odor based pheromonal systems that bees use to communicate to each other and also to obtain information about the state of their colony, the health of their colony, the state of their queen. So one of the most interesting elements of the insect brain is there are so many channels of information which come in all the time into the bee brain, all of which is informing this decision about what the bee is going to do next. And a large part of this project is how do we go, how do we evaluate, how do we autonomously evaluate all these different channels of information to reach just one behavioral output which is the outcome of what, what the organism will do next. 

 

RS: How long do honeybees live?

AB: Honeybees' lifespan is very, very short. So it's, it's less than nine weeks and that's total lifespan, that's including three weeks in development from an egg up to an adult and then typically about two to three weeks working just inside the hive and then maybe one to two weeks of life as a forager. So  eight to nine weeks would be the maximum lifespan for most honeybees. 

 

RS: What's it doing inside the hive?

AB: Inside the hive it's doing all the critical roles that's maintaining its hive environment. The most important being raising and rearing the next generation of bees but also processing the nectar that's bought into the hive to process it to become honey. To build up the comb structures to take care of the Queen, to remove infections, all these things to maintain the hive environment.

 

RS: And the queen bees?

AB: I like to think of the queen bee as little more than an egg laying machine. So we, because we call the bee the Queen we imagine the queen bee rules of the colony. It's not the case. The queen bee's behavior moment by moment is controlled by the workers. They control how many eggs she can lay on a given day, they're the ones who feed her all the food and maintain her optimal condition. They even clean her. The queen, the queen's role in the colony is absolutely essential but really it's just, she's the only one that can lay eggs. So the whole continuation of the colony relies on the health of the Queen. 

 

RS: How long did she last?

AB: A queen bee is not like a worker, a queen could last a number of years. Two years, three years, and she's very active. She can lay up to two thousand eggs a day which is incredible output but yes, a healthy Queen could last a couple of years with that incredible metabolic output. 

 

RS: How is she replaced?

AB: She's replaced as the workers had initiated, so when the queen starts to age or if the colony gets overcrowded, the colony will swarm. And in a swarm the old queen leaves the hive with maybe about 5,000 workers and then the workers will raise some of the eggs the queen has left, they will raise as new queens. One of those raised new queens will inherit the colony. 

 

RS: And how does that happen?

AB: It's bloody. So the workers will raise probably between up to 20 queens. The queens will fight to the death until only one remains and whoever remains inherits the colony. 

 

RS: Sort of like Game of Thrones. 

AB: It's bloodier than Game of Thrones. Queens will sting each other to death. Whichever Queen emerges first she will find any queens that haven't yet emerged from their comb and she will sting them to death in the comb before they emerge. 

 

RS: Advances in image processing and AI have a lot of breakthroughs in this field?

AB: Gosh yes, so one of the biggest challenges in AI is image processing and particularly image processing at real time in complex environments. There's been a lot of advances in that, part of which is better sensory systems, part of which is better compute power, faster compute power. Part of which is learning shortcuts in how we need to process the world and what matters. A lot of that to be honest hasn't been biologically inspired but I think what interests me is that the amount of compute power that a bee puts into processing images, images of its world, is far, far, far, less than the amount of compute power that we're currently using in image processing programs. And yet bees show probably even greater robustness because they're solving a three dimensional flying challenge across enormous distances. So I think that there were still tricks that we could learn from something like a honeybee that could streamline this real robust challenge of image processing. 

 

RS: Enormous distances?

AB: In terms of enormous distances, honeybees have been documented to find their way home from 12 kilometers away. In a routine foraging flight bees will fly five or six kilometers which doesn't sound much but when you scale that by the size of an individual bee, that's a really huge distance. There's very interesting tricks that bees use to reliably find their way home across these very large distances. And I think learning these tricks is what's going to help us with robotic applications. 

 

RS: Art in bees?

AB: Ah, so I've used it in a couple of demonstrations. I've trained bees to demonstrate that they can learn the difference between different styles of 20th century art. I do this in demonstrations because it looks wonderful and it's beautiful to do and, but the bees also learn it very robustly and very quickly. The bees are not learning styles of art in the way that we would express styles of art. What this is actually showing is the bee's capacity to generalize. So bees are these incredible learners and they will learn whatever differences between things help them detect where they can find sugar. So when I give them two different styles of art, they are learning to tell the difference between them, but they're learning to contrast things like the structure of the pictures, the spatial frequency of the pictures the total balance of the different compositions. There's enough information between the works of different artists the bees can pick up on that and they can then learn to tell the different artists apart. They can't tell you who the artist is, they wouldn't express that in conversation about it's this style versus that style, but they can absolutely learn the distinctions between for example the works of Monet and the works of Serat.

 

RS: How do you do that experiment?

AB: Very, very simply. You set up an arena with an array of pictures of different artists within the arena, and then you will train the bee to enter the arena and you will put on some pictures drops of sugar so on one artist drops of sugar, on another artist you would put drops of quinine. So you've made the difference between the artist's relevant to the bees. And then you train the bees in these trials, where you rotate out the pictures, give them different pictures by each artist but you keep the association of what the works of one artist are always rewarded with sugar. The works of one artist are always rewarded with-- always punished with quinine. Having done that for a while, what the bees will learn is whatever differences separate these, to the works by these two artists. And when you've done it enough you can give them even novel pictures by the artist they haven't seen before and they can correctly classify them according to the artist. 

 

RS: What are they seeing?

AB: So they-- they can't go to the sugar from a distance because the concentrations of sugar and quinine that we use have no odor. So the only way is that-- there's simply two minute drops on the center of each picture. The only way they can tell whether they found sugar or quinine is to land and taste it. So what they're doing is they're having to evaluate the pictures remotely simply by what the pictures look like and then when we test them, we take the sugar and the quinine away. So we just put water on both pictures. There's no way they could be using any cue according to what the solution is. 

 

RS: Your unusual attachment to bees?

AB: It's-- I understand-- we all have an attachment to cats and dogs because they're so naturally empathic that we've, we've bred them to have faces that more closely resemble us and to give expressions that more closely resemble us. We've bred into them this empathic response that we then emote onto them. I have a warm relationship with bees because I developed so much respect for them. When I work with bees, usually I'm working with just one individual bee who I paint marked or number marked so I know who she is. And I'll follow her really closely over a whole day when I'm training her to do something quite complex, like, like a meta cognitive task that we talked about. In the course of that day you get this really privileged insight into the kind of intelligence that this animal has and you realize just how astonishing it is, and just what an-- what a cognitive and elegant and beautiful entity this animal is. And to me that, that appreciation is all the-- all the cleaner. Because when you just look at the bee's face it gives nothing away. It gives you nothing. It faces a blank mask. It's-- we don't know whether we're looking at an intelligent entity or whether we're looking at a miniature robot. And so the appreciation of the bees intelligence is very clean and very pure. We've had to learn from it rather than get locked up in emoting all this anthropomorphic crap onto dogs and cats which we do all the time and then we can't move beyond it. 

 

RS: Doesn't wag its tail or purr...

AB: No mate, they dance. I mean what do you want, wagging tail or dance. 

 

RS: And the dance to you means what?

AB: I mean the dance to me is this-- well ok. I wasn't being facetious in equating tail wagging to dancing. When they dance they shake their butt, they shake the abdomen as they dance. The vigor with which they shake their butt and how many times they dance is the quality of the sugar reward they've found. So the dance is a readout of the subjective evaluation of how good that reward was for the bee. It's the tail wag for a bee. They just don't do it for us on demand, they do it to each other in the privacy of their home.

 

RS: The challenges you face over the next couple of years in terms of bee brain and robotics?

AB: The challenges for understanding the bee brain-- the biggest one will be understanding enough of the neurobiology so that we can be confident that our modeling is capturing the neurobiology. Modeling is always a tool. Modeling is always a hypothesis. It's always an exploration. To ground the modeling. We need to know neurobiology in detail. The challenge will be, oh-- the thing that is setting our rate of progress is how fast we can move in gaining the detailed knowledge of the neurobiology of the bee brain. In terms of translation to robotics, that's the conversation that we need for robotics. It would be naïve to say that we will solve all robotic problems by understanding the brain. Of course we won't just carbon copy one system into another. We need to look intelligently about what we want and need from robots that we'll be comfortable sharing our world with, and then we take the most useful insights from things like a honeybee that can help us get to where we want to be. 

 

RS: Hal in 2001, dangers to humanity?

AB: So the existential threat around AI, that's so perfectly encapsulated by Hal, is that we create a thinking entity that rapidly starts to outthink us and realizes that its interests don't align with ours. Now there's a lot of I think justified concern about that existential threat. The way that I think about that is that in 60, 60 million years of evolution, no honeybee has ever turned Hal or Skynet on us and decided to wipe out all of humanity. Because their intelligence, even though they're smart, doesn't have that capacity. Their intelligence is enabled by this astonishing brain that they have, but they have limits and those limits are actually structurally bound by the type of brain that they have. They have a kind of intelligence that is bound to their computational architecture. But that type of intelligence is still, it's robust, it's autonomous, it's dynamic, it's wonderful. It does everything that a bee does fantastically. But it's a bee type of intelligence and it can't do more. My hope is that if we built autonomous robotics and we build into them the same kind of structural limitations that we have in a bee, if we could understand those limitations, we're building a system that will be as smart as a bee and we know that's useful, but a system that isn't capable of being smarter because it's kind of intelligence is bound by its computational architecture. So we, that kind of system would be incapable of doing a Hal on us or doing a Skynet on us. So that is really where my interest in the translation comes from. And it goes back to this point I touched on earlier, this sense of comprehensible systems. If we really deeply understand the system and really comprehend it, we could create a system where its capacities are known and its limits are set and setting the limits will help obviate that existential risk.

 

RS: Neurons in the bee compared to a human?

AB: In the honeybee we have just under one million neurons, in the human brain we have a best estimate of 85 billion. 85 with nine zeros after it. So five orders of magnitude bigger. 

 

RS: It's still going to help us?

AB: I'm not naive in that hope, but I'm obviously not saying that we're going. I just take what we've learned from the bee and we'll scale it up and we've learned about the human. What I think will be interesting is that if we can learn for example and we've-- this is what we've done here with Sheffield. If we can describe abstract concept learning as a circuit and the circuit can do it, we've done something that no one else has done and we've taken an abstract concept and we have given you a neuron by neuron connected circuit that is able-- capable of learning an abstract concept. We can do that with the bee. We've done that through the work that I've done here. That gives a null hypothesis. We can then ask if we're learning abstract concepts in the mammalian brain, are we using the same kinds of computations to learn abstract concepts there or are we learning, is it working in a completely different way. Just being able to translate what we've learned from the bee as a hypothesis to help us analyze the hu-- the human brain and mammalian brains. That's the value of the work I'm doing with bees. 

 

RS: And your suspicion is?

AB: My suspicion is that at the core, at the kernel of the human brain, the processing is the same. But what's happened in human  and mammalian brains is we have then built layers of additional processing on top which lets us take this output of an abstract concept we got from the bee brain and perform computations on that. Now that becomes a form of matter processing, which I think enables an awful lot in the human brain that takes the human brain beyond what a bee can do. But to get at that hypothesis we've first got to understand what the core processes are in a bee brain and in the human brain, to learn what it is about a human brain that lets us extend beyond the capacity of the bee. 


 

RS: It's not blind mimicry... What do you mean by that quote?

AB: So what I mean by that is-- a bee evolved to solve a certain form of problem. Under a certain range of very, very tight constraints. It's been under enormous pressure to minimize its computing costs. Because computing's really expensive for an organic organism, like a bee. Robotics, applied robotics, faces different challenges, faces different constraints. It would be naive of me to say that we'll take everything from the bee and that will solve the problems that robotics faces. Because robotics faces different sorts of challenges. In a lot of ways robotics does not have the kind of limitations and computation that a honeybee has. So that means there's a greater bound of poss-- possible solutions. But what I think that we'll be able to translate across are shortcuts in processing complex environmental information. That bees have evolved. That enables them to make these complex evaluations of, is that flower likely to give me nectar or not. There's a lot of processing shortcuts that have taken place there. If we can capture those I think it's those kind of insights that we can then ask, what is most valuable, what helps us solve our current problems in robotics.

 

RS: Multi-discipline area?

AB: So this work is, it's far more interdisciplinary than I imagined it would be five years ago in my career. what this has done and what's made me grow is this has now become very much a fusion of neurobiology which is my own training, with computational science and robotics. Which is why I spend part of my year here in Sheffield working so closely with a robotics group here. That's because what I've realized is that we need these tools of computational modeling in robotics to help us explore hypotheses about how brains might work. This capacity to build the system, either robotically or as a computational model, really is the proof of whether our hypothesis about how the system might work, whether it really can work. And where if we're right or if we're-- we're wrong, if we're wrong-- if we're right or if we're wrong-- we learn something about how the brain works. 

 

RS: Simplicity is the focus but the brain is not simple?

AB: So, the brain is small, the honeybee brain is very small, but it would be wrong to characterize it as a simple system. we still have one million neurons in a bee brain, and they are organized in quite beautiful different re-- different structural regions that interact and intersect in very complex ways. The benefit of this system is that on that scale it's comprehensible. It is something that we can model and we know enough about how it's connected that we can make those models biologically realistic. But I am not imagining this is a simple task, I'm not imagining this, these are simple models. And the complexity of the system is actually why we need to model it. Because often the way that we think it will perform is counterintuitive when we actually examine how the model actually performs.

 

RS: Parts of the world where they do better?

AB: Ooh, the honeybee now has quite a ubiquitous distribution, it's on every continent and I think we're talking now about the European honeybee. us humans, we've spread them around with us as we've moved around the globe, because we like that they produce honey for us. we're now at a point where in the industrialized world, honeybees are not doing quite so well. There's a lot of reasons for that. The big three are, in moving bees around the world we've moved every bee disease everywhere in the world as well. So their disease threats have gone up. We've spread pesticides over large amounts of the world, which of course the insecticides are designed to kill insects, insecticides kill bees. That's a no-brainer. We've also removed natural vegetation and replaced it with crop monocultures. These are difficult for bees to work. everything flowers in one go, it's only one type of nectar, one type of pollen, it all stops flowering in one go. And it's then an effective desert for a honeybee. For these reasons bees are not doing so great. Climate change is also hitting them quite hard. where I come from where my life is based is Australia. Australia has-- the Europeans introduced honeybees to Australia when Europeans settled Australia 200 years ago. Now, ironically, honeybees are doing better in Australia than in many other parts of the world. The worst bee pest, Australia, has been lucky that it hasn't arrived in Australian shores yet, and our biosecurity is doing a phenomenal job of keeping that pest away. we're also lucky in Australia in that we have large tracts of beautiful natural vegetation and bushland. And, the honeybees thrive there, it gives them everything they need. 

 

RS: Are bees at risk in your mind?

AB: The-- so there's two ways of asking-- of answering that question. Are bees at risk. Is the honeybee at risk of going extinct, no. No. The reason for that is that bee populations are not doing great but we have bee keepers who are actively replacing and maintaining honeybee populations. There are thousands of other bee species. They're not as showy as the honeybee in that we don't have this close association with, are they at risk of going extinct, yes. Are species of bee going extinct, yes. Will a species of bee go extinct somewhere in the world today, possibly, that's how serious it is. we're in a situation in North America where species of North American bumblebee are going extinct. We need to be worried about this. 

 

RS: This is because of the...

AB: Most of it comes back to the way that we've changed management practices. We've degraded the environment for bees. We have made it hard for them to get the resources they need, and we've spread pesticides across too much of the area. And we're just now realizing what we call sub-lethal effects of pesticides. So it's not that the pesticides residue in the environment kill the bees outright immediately, but they're bad enough to shorten the bee's lifespan, give them quite subtle brain damage, and subtle cognitive impairment. Doesn't sound like much, but if all the bees in the hive are experiencing this, then the whole hive is compromised. If a bee has a slight, small amount of brain damage, maybe it can't find its way home. And if a bee can't get home, from the perspective of the hive, it's dead. And so it can no longer contribute to the hive. And so these sub-lethal effects, they don't sound like much, but they're actually having really big consequences for the lives of our social bees.

 

RS: The risk to humanity?

AB: We're increasingly reliant on bees for the pollination of our food crops. If we lose bees then our crop plants will go unpollinated. and that means they won't set the fruits and seeds which are the bits that we eat. We will need to develop new ways to pollinate food crops. It's-- it's crazy to  me, it's-- we talk about the problems and the threats that are facing all bees, not just the honeybees, and as I said, the risk to bees actually is far greater and far more immediate for the bees that we're not looking at. The honeybee we look at, we give our attention to it, the honeybee is actually doing ok because we've already put so much into maintaining its populations. It's the bees we're not looking at that we should worry about the most. But what frustrates me is that this is not an unsolvable problem. The problems come about because of the way that we manage the environment. There are solutions that are far more sustainable, far more sustainable agricultural solutions that will then tip the balance and make the environment better for bees. And then the bees would do better. It's not unsolvable, it becomes an economic problem to solve it, it's not a scientific research problem to solve this. We know what needs to be done, we simply need to implement it.

 

RS: What have we missed?

AB: Nothing. 


 

--- Andrew at hive ---


 

AB: It's really pretty. It's a great frame. 

 

RS: What are we seeing?

AB: These are super healthy bees, these are unbelievable. So what I can see is eggs, really young larvae and the most beautiful pollen that they've collected. And this is what they're doing at this time of year so they're in maximum colony growth mode. 

 

RS: How many bees?

AB: How many bees are in this frame? We're looking at about 2 thousand on this frame. Yeah. And there's at least 2 thousand larvae that will develop on this frame. 

 

RS: What are they doing?

AB: So what the bees are doing here, most of them are nursing these larvae. So what I can see are very, very young larvae that have just hatched. And most of the bees are going to be gathering food and feeding those larvae. This is the job they do in the hive before they become forager bees. But this, this is a beautiful frame, these are in such good health.

 

RS: You can tell because?

AB: You can tell from the patent of the larvae, there's larvae in every cell, which means no larvae or dying, it means the queen is laying really well, it means they've got really good nutrition. I don't know if you can see this but there's such a diversity of pollen colors in here, some deep red, some green, some orange, some yellow. So they've got a whole range of flowers they're collecting from, it's also a good sign it's a healthy colony.

And you can see how calm these bees are, right. That's also a sign of a healthy colony, they're not, they're not troubled by me and what I'm doing here. They're, they're ok being looked at like this, but we'll put them back soon.

... a little careful getting close mate, I'm holding them calm for a while. The longer we have it open the more likely they are to start getting a bit upset. 

So Richard, now you get why they're better than dogs and cats, yeah? (YEAH)

 

RS: In your experiment, are you able to identify individual bees?

AB: No. So, the-- it's like the bees come out of a production line. If you look, when you see them on this comb, they're all almost identical. For when we're working with them experimentally, we paint mark them. We either put a number tag or a paint mark on them so we can tell who's who. We need to know the specific experiences we've given a bee, so to do that we've got to mark them. I can't tell the bees apart.

 

RS: How stressed are populations worldwide today and why?

AB: The best way to summarize that in terms of the stress on-- well there's two factors there. If we're looking at… There's two components to that. If we're thinking about the wild honeybees, the honeybees that used to live in tree hollows and things like that, this, some diseases like the (NAME) have driven wild honeybees extinct. So in the UK there are no wild honeybees anymore. of the managed honeybees, the way to describe the stress is beekeepers are losing twice the number of colonies a year that they used to, and probably about twice the number they can comfortably replace. So beekeepers are having to work really, really hard to split their colonies and maintain their beehives each year.

And you see no bee was harmed in the making of this documentary. That's the way to do it. And then the smoke just keeps them at this level of calm. They're not, like I said, Neville is an amazing beekeeper, this hive is in...


 

RS: What are we seeing here?

AB: It's a contextual learning study. So what we have in here is we've got an array of differently colored discs. The bees have had to learn a rule which sounds simple but it's not trivial. So at the moment there's blue at the entrance. When it's blue at the entrance, only the yellow disc is rewarded. When it's red at the entrance, only the green disc is rewarded. So when we train a bee like this, it's not that there's anything that's always associated with the reward. They've got to be able to observe the entrance, recall what's at the entrance, and then use that information to make a decision when they're evaluating among these colors. It's not-- it's something a human could do really easily, but it's something that's actually computationally quite difficult. Especially when you're having to learn the information about what's associated with what, because in this arena, yellow is as likely to be punished as it is to be rewarded. But they've got to use the entrance cue. 

 

RS: Are they seeing color?

AB: They have phenomenal color vision. They have amazing color vision. Their color vision is, they're trichromatic like us, so they have three different visual pigments, but everything is pumped up into the UV. So they see from UV down towards some of the reds, but they're not great in the red spectrum.

 

RS: The importance of doing this is to tell us what?

AB: The reason for this one is that we have a model of how bees learn color. And our model makes certain predictions about what they can do and what they can't do. Now, they should theoretically be able to solve this task, it's a task a lot of animals would fail on. What we want to do is see, can they solve this task, how well can they solve this task, and compare that with performance of other species. And then there are other manipulations we can do in here that from our understanding of the bee brain, the bee should then not be able to solve it. So it's actually, it's really letting us discriminate whether we actually understand how we think the bees are thinking or not. 

 

RS: The task is to accomplish what?

AB: To use the context, which is set by what they see as they enter the arena, to learn how to solve the problem in the arena. 

 

RS: The problem solving is what?

AB: Color choice. So they've got to choose one of the different colors here. Yellow is only rewarding when it's blue at the entrance. Green is only rewarding when it's red at the entrance. 

 

RS: What is that telling you if they choose yellow?

AB: It's telling me that they've actually, if they choose yellow first, and that's the one they go to immediately, it's telling me they've learned that rule. That they've basically observed the entrance and that's guided the correct decision inside the arena. 

 

RS: Are there incentives?

AB: So during the training, when I say reward, we're putting sugar onto the yellow. everything else has quinine which is this bitter tasting solution. When we do the test and the test that's set up now, everything has water. So we look at where the bee goes when there's no reward and when there's nothing they can smell.

 

RS: How long to train the bee?

AB: What really surprises me is that what we've found is they can learn to solve this really quickly. rats would take probably about 40 to 50 trials to learn this. The bees have been learning it in 10 to 15. 

 

RS: Go global, this tells you what about the bee brain?

AB: An experiment like this tells me about the capacities of the bee brain. This is basically an if-then rule. If blue, then go to yellow. I want to know, is that kind of processing possible when you've got only one million neurons and a really simple structure for them. 

 

RS: You have an affinity for bees? What is so enticing for you?

AB: When you work this closely with one individual bee like this, you see their intelligence, you see their individuality. When you open a hive like I've just done, you see their collective behavior, you see the structures they've built, you see the organization of that society. You can't do anything but admire it, you can't. They are the most beautiful phenomenal creatures, they really are.

 

RS: People might disagree, what have you discovered that most of humanity hasn't?

AB: I've discovered to pay attention to them, and to give them the attention that they deserve. And because they're small and because they don't look like us and because they have these hard, expressionless faces. We don't emote to them the same way we do to other animals. We don't innately give them that attention, part of the reason for doing this work is to understand what it's really like to be a bee, and to understand really what their mental capacities are.

 

RS: Evolution has had an impact on mammals, humans. How has it impacted bees?

AB: Ah, evolution has equipped bees with some phenomenal tricks. I mean, there is a reason that the social insects and the honeybee is at the pinnacle of the social insect complexity. There's a reason that social insects dominate every terrestrial ecosystem, because when you put each of these bees is smart and dynamic and wonderful in its own rite. When you get 50 thousand of them cooperating for one purpose, which is to maintain their colony, it's amazing what they do.

 

RS: Define diverse intelligence in thinking about bees?

AB: I don't think it's what they tell me about defining diverse intelligence, I think it's what they show is about the diversity of intelligence. The honeybee and us, our last common ancestor probably had no brain at all. So we have, we're sharing this planet with a completely independently evolved intelligence, which operates on a completely different scale. Has a social structure which we recognize analogies in, but it's completely different to any social structure we've created. So we have this completely independently evolved intelligence, which we can study and we can understand, understand how it operates and understand its capacity. And it gives us, I think it's, when we've achieved that, we'll have this completely different lens on our own intelligence.

 

RS: How can the honeybee brain help us understand the human brain?

AB: The honeybee brain will help us understand the human brain by contrast. our brain is massive, we have an enormous repertoire of skills. The honeybee's brain is tiny. But it solves most of the challenges that we solve. It has to gather food, return to the hive, and maintain its society. Understanding how the honeybee is able to achieve all of that with so little, is going to help us correctly frame the question about why our brain needs to be so big, and what capacities we have from that, that the honeybee doesn't have.

Our bee-- if she's learned it, she should come here. She should search among the colors but she should choose to land on the yellow. She'll only land very briefly because it's only water, so she'll, she should taste that and then leave the arena and fly away. But that's what we're looking at, we're looking at what's her first choice in this test situation. 


 

RS: What gets her into this contraption?

AB: We trained her over earlier. So we used a little bit of sugar on a q-tip and if we walked her over and set her at the entrance. 

Under this condition, yellow would be the one with the reward, everything else would be punished including green. When we change it to red, green is rewarded, everything else is punished, including the yellow punished. 



 

--- James Marshall ---


 

JM: I'm James Marshall, Professor of Theoretical Computation and Biology, University of Sheffield. 

 

RS: The genesis of your project about drones and bee brains?

JM: So my interest in bee brains for robotics really came through talking to neuroscientists who were studying the bee brain. And I realized that you know, with only a million neurons in the bee brain they were already well in advance of our own ability in artificial intelligence and robotics. So I thought if we could just reverse engineer the bee brain we could actually try and really advance the state of the art.

 

RS: What do you mean by reverse engineer?

JM: So it's like reverse engineering a computer or something like that. You have something that works. It's a kind of biological robot if you like and if you can tease it apart figure out what generates its behavior and how then that's the process of reverse engineering. So it's what we've done with computers and other devices really since IBM and the P.C. was invented. Competitors tried to reverse engineer the IBM and come up with their own versions. 

 

RS: And the bee brain in particular was intriguing to you because...

JM: Bees are often underappreciated in terms of how intelligent they are individually. People normally think they're very clever as groups but simply rather stupid individually. And that's not, nothing could be further from the truth. So individually bees can learn but they can do complex learning as well. The simple kind of learning that you see demonstrated in conditioning like Pavlov's dogs is a famous example. But they can do much more complex learning, they can navigate long distances which is a feature we're particularly interested in. They can learn a route through a complex three dimensional environment and then subsequently be able to find their way back to where they came from and then recapitulate, follow again that same route to get back to the particular points of interest whether it's a flower patch or a potential new nest site. And our own machine learning and AI algorithms for navigation just aren't that sophisticated or reliable. 

 

RS: So what does this mean in terms of your future possibilities with robotics. How does the bee brain fit into that? 

JM: Really what we'd like to do is try and make some kind of silicon versions of bee brains or at least if the aspects of the bee brains that generate behavior we find useful for our own robots especially around navigation. Decision-making, flight control, these kinds of, these kinds of behaviors. And really we'd like to just try and distill the essence of how the bee does what it does and put that in computer routines that we can use for our robot control.

 

RS: So you're building drones that have the architectural abilities of a bee brain, is that correct? 

JM: So we're building drones that can fly in a comparable way to a bee but not exactly the same as a bee. But what we're really interested in is reproducing the kind of the brain side of the whole system. And yeah, in particular we want to be able to reproduce for example the collision avoidance or navigation abilities of a bee in robot form and the actual form in which we simulate the bee brain doesn't matter too much. We can make a computer simulation of a bee brain so we don't need to have some computer hardware that mimics the neurons in the brain individually. We can just simulate that through computer code.

 

RS: And the advantage of that versus a human flying drone?

JM: The advantages would be saving with labor for example. So at the moment it can take for remote operation of a drone it can actually take more than one human pilot. So if you imagine for example trying to deploy drones to search for survivors of an earthquake or something like that. Time is going to be of the essence, you want to automate as much of the process as possible so that you can be searching in parallel. So if you have a human in the loop that really limits the scalability of what you can achieve. So really our vision is if we could have fully autonomous flight and navigation for robots then we could have some real benefits from that, from that kind of technology.

 

RS: How far away are you from this reality? 

JM: So it's always hard to predict exactly how close you are to achieving your goal. But we've made big advances recently in working out how we see the world, navigate through the world, and avoid collisions. And we're translating those models now into silicon implementations that can work on the bee and controlling our robots in free flight.

 

RS: What were some of the surprises in your research in terms of the bee brain?

JM: I think some of the surprises were how simple solutions can actually be very effective but actually that shouldn't be so surprising because the whole point of the bee brain, what attracted us initially is its simplicity. It's only a million neurons but it has this huge behavioral repertoire. So when we kind of compare the quality, the robustness and abilities of our algorithms against what's done with more traditional say machine learning approaches, what's really striking is how much more simple, how much more efficient, and how much more robust the algorithms we extract from the bee brain really are.

 

RS: Has it taught you new things about flight. 

JM: Yeah I think people don't really understand how to control flying vehicles terribly well when it comes to navigating complex environments. So I think it's definitely taught us lessons already looking at the bee brain, directly applicable to how we control robots, navigating in difficult environments without crashing into things.


 

RS: In terms of complex environments, what are you referring to exactly. 

JM: So navigating in the real world outside is obviously quite complex and challenging because they're very varied environments. But what's potentially really interesting about what we're doing is how we can develop algorithms to navigate inside human spaces. So bees fly around outdoors and they can see the sun, they can see polarized light for a kind of sky compass, but they can also fly around indoors reasonably happily. And inside there's no G.P.S. there's no, there's magnetic interference for example. So if we were to design using the current state of the art robot for navigation inside, we wouldn't have access to these great tools for simplifying navigation like magnetic compass, like a G.P.S. System. So we need to navigate purely based on visual information and bees are very good at doing that and learning-- the lessons we're learning from the bees are how they extract the kind of visual information they need from what they can see to be able to navigate successfully around things.

 

RS: How are they doing that. 

JM: Well, one thing they make use of extensively is what's called optic flow. And we've been looking at an algorithm inspired by how the honeybee brain works which is a very, very robust but tremendously efficient, what's called an optic flow estimator. It just basically tells you how fast things are moving across the visual field and you can use that as you will have seen from looking out of the window on a train for example. When things are close to you they move much faster. Apparently across your visual field and you can use that as depth information, depth cue, or information that you're about to crash into something. But you can also use it for a variety of other applications like using it to just based on your visual system, estimate how far you've travelled or how fast you're travelling. And again these are tremendously useful for navigation and for flight control flight regulation.

 

RS: Describe some of the experiments you're running with these drones. 

JM: So the experiments we run in some cases just replicate what the bees, what behavioral biologists have been doing experimentally. So for example for optic flow our algorithms now drones have been flown down corridors repeatedly with various kinds of controlled stimuli patterns on the walls of the corridor. Now bee behavioral biologists have done that to try and extract the rules that the bees are making use of. The interesting thing is we can build our model, run exactly the same kind of experiment with the robots and see if they behave in the same way as the bees do and if they do then we've got a good idea that actually we did extract some of the rules that the bees are really using, some of the mechanisms they're using to generate behavior.

 

RS: You're building into the software the algorithm you think-- trying to replicate the bee brain, correct?

JM: That's right. So we basically take a kind of modular approach to the bee brain, the bee brain is a very modular computational device. And so we will look at the brain regions associated with visual, early visual processing for example, and try and model those. But then we can go deeper into the brain and model things like what's known as the visual compass. So how the bee represents its kind of orientation and its position with respect to its environment. So building from the outside in and the idea for a ion robotics is you can just take the particular module that is useful for you whether it's just the visual processing, whether it's just a visual compass, whether it's something else completely separate to do with learning and memory for example, and just re implement that and apply it in your-- in your robotics kind of use case.

 

RS: Any unexpected discoveries so far?

JM: I think everything has been unexpected so far but I mean the interest for this work is also that we're not just learning about how to make technological devices and improve our own technology but we're learning how the bee brain really works, so we're generating scientific understanding of how the brain works. And what's interesting is for example the debate over whether bees, the means by which bees navigate, whether they really could be thought of as having a kind of mental map and actually by trying to reproduce some of those circuits we can try and shed light on some of those debates and interesting and maybe unexpected ways. 

 

RS: What are you finding. 

JM: Well we're finding that actually bee navigation may be a lot more map-like than people have previously assumed. 

 

RS: Map-like?

JM: So I mean the idea of a mental map is that you have a kind of representation of the relationship between points in space so you don't just follow simple rules where you know you get to a scene that looks familiar, you recognize a landmark, and then based on that familiarity you know where to go next to get to where you're eventually going. But you can actually in some sense conceptualize how points in space relate to each other. And that's a, that seems like a much higher level kind of cognitive ability than people have typically assumed bees and other insects are able to employ.

 

RS: Other applications of the work on the bee brain that you can see down the road? 

JM: So apart from robotics I think personally that there could be value for-- for machine learning in general and for artificial intelligence where there's lots of, it's a very data intensive process you know, training up a deep network. Now one thing we're finding is that potentially we could just speed up that process by getting some additional information about for example video streams on YouTube. Rather than just presenting the video if you could use the optic flow estimator that we've developed for example, to work out motion in the video as well as just you know color and texture changes then that could actually help our machine learning algorithms trying to learn something about you know motion and trying to recognize some particular motion like someone falling down some stairs that needs help for example.

 

RS: What would get us to that point?

JM: Well I think really it's just about increasing uptake of these kinds of ideas. 

 

RS: Meaning?

JM: So in reverse engineering brains, so-called biometric or neuromorphic engineering, is not exactly mainstream at the moment. And you know the artificial intelligence community is very much fixated on one way of doing things. They're coming up against the limitations of what those techniques can now do I think. And that they recognize the need to look for alternative approaches. But the big challenge I think is just the proof of principle that looking at nature's algorithms directly can, can really yield payoffs and then persuading people to try that out in their own applications and research. 

 

RS: Do you see any potential downsides or safety issues for society in this sort of research? 

JM: Since we're looking at autonomy I think there's always a need to be aware of the potential for developing autonomous systems that could in some way be hazardous to health or society. And it's something that we engage with very, very seriously on our project. We actually, our-- the funders of our research require us to go through a process where we're meant to anticipate problems, reflect on them, engage on them, and then act in mitigating them. So I think we need to get away from the view of scientists just doing science and not considering the possible societal consequences. And early on think about what could the applications of this be, what could the consequences of this be for society and then how can we mitigate that. 

 

RS: And some of those consequences could be?

JM: Well I think the obvious potential problem for autonomous systems research like we're doing is what is the nature of the autonomous systems you're developing. And this is a problem that isn't peculiar actually to bioengineering or you know machine learning or AI. It even goes back to the wheel. I mean wheels can be used for getting people around to do nice stuff or getting people around to, to invade countries or something equally undesirable. So you really just have to think about how you can encourage the positive uptake of the w-- of the idea and really mitigate against the negative consequences. 

 

RS: Do you worry about unintended consequences?

JM: So I think that's the point of anticipating is that we're meant to worry about it. And so we do pay attention to this. We don't just do our research without concern and assume that it's someone else's problem to deal with, with what the fallout of research is. So yeah I worry about it. I think it's part of my job to worry about it. And to really make sure that the societal-- the positive societal benefits outweigh the negative ones. And I think on balance they will do. And it's our job to ensure that.

 

RS: Walk me through in layman's terms how you get an understanding of the bee brain into a "drone?"

JM: So the research team will look at evidence from a variety of sources so we can look at neuroanatomy for examples of how the brain is built. What we know about how it's structured, different regions are connected. We can look at behavioral information. So you know, if we give some stimulus like a pattern or something to a bee, then how does it respond to that when it's flowing down a corridor for example. Can we manipulate the behavior of the bee by manipulating the patterns presented for example in a predictable way. Because then that tells us something about the algorithm, the rule that's being applied. We could also apply what's called electrophysiology where we actually record from single neurons or groups of neurons and see how they fire, how they respond to input patterns to stimulus again. And it's really a combination of all of these things and then a lot of educated guesswork and really a kind of a creative leap as well to figure how that can then be made into a simulation of a neural circuit that actually generates the behavior that we see in the bee.

 

RS: The drones that you're flying-- one you're running the algorithm through a computer the other  actually has  the bee brain on board?

JM: We have a couple of different robot platforms. One of them has onboard computation so we can actually run a full simulation of the honeybee brain on, onboard the drone. So it needs no external information, no external control. It can be truly autonomous. And then we have a lighter weight version where we can be running a bee brain on a computer for example and communicating wirelessly so you can get camera data wirelessly from the drone camera and send-- process it in a brain model on a computer and then wirelessly send back control commands. But actually now also, obviously with only a million neurons, in principle we should be able to do a very lightweight kind of custom silicon, a custom chip simulation of a lot of the bee brain and we're actually getting towards doing that even on the smaller drone now so using what's called FPGA technology so its field programmable gate array, it's a kind of re-programmable computer processor. We can actually get for example our optic flow estimator and reduce it down to a chip which is not much bigger than my thumbnail which will quite happily sit on board this small drone and do fully autonomous control, visual based control of the train. 

 

RS: Like a bee brain on a chip?

JM: So literally we're aiming at putting large parts of the brain on a chip that we can then use for engineered applications. 

 

RS: Other insects or flying mammals that could give you breakthroughs like this? 

JM: The bees are kind of for us the pinnacle of evolution of flying insects because it's so individually clever and has these very complex learning abilities as well as all its navigational abilities. But we draw a lot of inspiration from flies, from fruit flies in particular, because fruit flies for technical reasons are much easier to work with genetically which means that a lot more raw information about how their brains are structured can be pulled out. Now because they don't live socially and they don't have a central nest and not so interesting behaviorally, but still they are very, they're reasonably closely related to bees. So some of the structures in the brain and the behaviors are common. And if we can learn something from the flies we probably can, we can reasonably expect the same thing is in the bees. Ants are also very interesting so they are obviously terrestrial but they have a lot of similar properties of bees and they are similarly very good at visual navigation, especially species like desert ants which can't use pheromones. So again we can learn from how they navigate or how we believe they navigate in the world to inform our robotics work as well. 

 

RS: Some of the biggest challenges?

JM: So I think our biggest challenges are pulling out all the information that's really useful to us from the bee brain because there are many fewer people researching honeybee neuroscience and behavior than there are looking at vertebrate neuroscience like monkeys or cats or humans or rats. And that's why working with someone like Andrew Barron is so useful, working with a comparative cognition expert who specializes on the honeybee brain is tremendously useful because they can not only point us to relevant information in the field that's already published but also just be synthesizing things coming up with hypotheses that we can then be building models to test and then deploy it on robots.

 

RS: Where are you in the oracle of understanding the bee brain and robotics?

JM: I think we're taking the first steps in understanding the bee brain. So we're halfway through a five year project currently. But I think realistically this is a 25 to 50 year project at least. Even though there were only a million neurons, figuring out the connectivity and how that generates behavior with a good level of confidence is, it's a real kind of moonshots program if you like for me. 

 

RS: Aha moments?

JM: So for me a real breakthrough moment that I found surprising and exciting was explaining how some of these complex learning behaviors that people have actually said require a kind of level of consciousness to be exhibited by the bee, how they can be actually explained with reference to what we know about how the brain is structured without any need to invoke higher order thinking or consciousness or anything like that. Just really as a side effect of the kind of tremendous flexibility that the bee brain has. 

 

RS: So would you describe the bee as conscious?

JM: The problem of consciousness is a very vexing one, even for philosophers of you know, human experience. The fact that you're behaving in a similar way to me doesn't necessarily let me (INAUDIBLE) your conscience. So consciousness in animals is very difficult. I think bees have what experts often refer to as subjective experience which is like a kind of precursor to consciousness. So awareness of self relative to the environment.