SCIENCE AT THE SHINE DOME canberra 2 - 4 may 2007

Symposium: Development and evolution of higher cognition in animals

Friday, 4 May 2007

Professor Mandyam Srinivasan
The Queensland Brain Institute, University of Queensland, Queensland

Mandyam Srinivasan holds an undergraduate degree in Electrical Engineering from Bangalore University, a Master's degree in Electronics from the Indian Institute of Science, a Ph.D. in Engineering and Applied Science from Yale University, and a D.Sc. in Neuroethology from the Australian National University. He is presently Professor of Visual Neuroscience at the Queensland Brain Institute of the University of Queensland. He is a Fellow of the Australian Academy of Science, of the Royal Society of London, and of the Academy of Sciences for the Developing World. Srinivasan's research focuses on the principles of visual processing, perception and cognition in simple natural systems, and on the application of these principles to machine vision and robotics.

Small brains, smart minds: Perception, learning and 'cognition' in honeybees

Our group deals with bees. Those tiny creatures that will make a smooth landing on the rim of a teacup or will fly several hundred metres in search of food, find a good source of food, and then literally make a beeline back home. And part of the work that we do in our lab is trying to understand what it is that makes these rather tiny brains, which weigh less than a tenth of a milligram, tick as well as they do. That is one aspect of low-level vision and navigation that, unfortunately, I won’t be talking about today because it is slightly outside the domain of today’s topic.

Another aspect is trying to see whether, if they do do something innate, we can put them to novel kinds of algorithms for controlling seeing machines and flying machines. That again is a little bit outside the area of today’s topic.

But the third area, which I think is the most exciting, has to do with how far you can push the perceptual capacities of a creature with a tiny brain. Lots of animals are thought to be behaving like automatons. And insects have classically been thought of as fairly rigid, stereotyped automatons. Even if an insect learns something, the tendency is to think that it will learn it in a fairly rigid, stereotyped way. The question we would like to ask is: is that really true, or can insects learn things in a more flexible way? Can an insect, for example, learn a particular thing in one context and apply it to a totally different context that it hasn’t encountered before?

So that is really the first question: are small brains capable of some form of generalisation?

(Click on image for a larger version)

We did this study quite some time ago. It had to do with pattern recognition in bees. It has been known for a while that bees will learn and memorise patterns and then associate a promising pattern with a good source of food. They will learn this pattern and come repeatedly to it, and will choose this pattern in preference to other patterns if the particular pattern happens to be associated with food.

The classical notion was that when they learned these patterns, they learned them in a fairly photographic, eidetic way, as in memorising a snapshot. Then, when they encountered a new pattern, they would simply take the old memorised pattern, see how well it matched – literally in a pixel for pixel way – with the pattern that was being seen. If there was a match they would say, ‘Yes, this is the right pattern,’ and if there was no match they would say it was the wrong pattern. We were asking ourselves: is that the only way the system works, or can they generalise patterns? Can they learn general attributes of these patterns and apply these to select between novel classes of patterns?

So here is the experiment. We train bees to come into a Y maze (‘a’) and choose between two kinds of patterns, in which the stripes in one are always vertically oriented and the other always has horizontally oriented stripes. At ‘a’, we reward the vertically oriented sets of stripes. The stripes are varied all the time – they are chosen from a random population – so there is very little opportunity for the bee to actually memorise each of these patterns in each trial, but it has to learn the concept of verticality and then approach the vertical pattern. And if it gets in there, it gets a reward. That kind of task they really can learn very well.

Similarly, in ‘b’ if you reward the horizontally oriented pattern, the bee can learn to choose that in preference to the vertically oriented pattern. Or if, in ‘c’, you go with patterns that are oriented +45° and -45°, they can do that as well.

So they can certainly learn a general concept of orientation and apply it to patterns that they haven’t seen before.

(Click on image for a larger version)

That is shown a little more clearly in this slide. We started with these random patterns at ‘a’, so the training is done on these random patterns that are oriented in one way or the other, rewarding on one orientation.

If we now test them with the novel patterns ‘b’, with just single stripes, they will pick the pattern that has the correct orientation.

Again, in ‘c’, which is a different kind of display, with a white stripe on a black background, negative contrast, they will again pick the stripe that has the right orientation.

In ‘d’ you have got thin stripes – again, the correct orientation.

In ‘e’ you have a single diameter on a circle – again, the correct orientation.

In ‘f’ we have a sinusoidal grating, which I haven’t shown in the slide but is actually a sin wave grating. Again you see the pattern that is being picked is the one that has the appropriate orientation, the one to which the bees were trained.

In ‘g’ we have patterns with bars which have a little gap in the middle. Again the correct orientation is preferred.

So clearly these bees are able to abstract this property of orientation and, we think, apply it to novel kinds of patterns that they haven’t seen before in their life.

(Click on image for a larger version)

Here is another interesting feature. The training is as before, with these random patterns with different orientations. You train them on the patterns at ‘a’ and then you test them on the rectangles, which have one orientation as opposed to the other. Bees that have been trained on the one orientation clearly prefer the rectangle in ‘b’ that is oriented in the ‘right’ direction.

But here is the interesting thing. If you show them the illusory pattern in ‘c’, which looks as though you have an illusory contour and an illusory edge defining the rectangle in negative form, rather like an envelope that is covering four circles, bees will now pick this pattern, suggesting that they really do perceive these illusory contours the way we do.

What is also interesting is that if you now break this illusion – for us humans, anyway – by rotating these patterns so that the illusion is no longer there for us, the bees also behave as though they can’t tell the difference between these two orientations any more. This is the closest you can get to asking a bee whether it sees the illusory contours the way we do or not.

What is interesting about this is that, cognitively, there is a high-level cognitive explanation for this and there is a lower-level explanation for this. The high level says that we humans look at this thing and say, ‘The most likely interpretation of this is that it is a white envelope sitting on top of four black circles, so that’s what it must be.’ So our brain fills in, in a ‘top-down’ fashion, the missing edges there. On the other hand, the low-level explanation is that the brain doesn’t bother about doing all that; it just looks at these edges and whenever there are two edges that are parallel and co-linear it just fills in the gap. That is probably what is happening in the case of the bee, because we don’t think a bee would actually have a concept of an envelope (it doesn’t deal with the post office) so it’s really a very low-level sort of a situation where it is trying to basically deal with occlusions – occlusions that happen all the time in nature – which the brain has to fill up, to actually pull out individual objects from their backgrounds.

(Click on image for a larger version)

Here is another thing which has never been looked at in lowly creatures. This has to do with learning mazes. As many of you probably know, maze learning has been studied ad nauseam in rats and mice, but surprisingly little work has been done with insects.

(Click on image for a larger version)

So we looked to see whether bees could be trained to fly through a maze. This is work done by my colleague, Zhang Shaowu.

We have a labyrinth-style maze here consisting of a bunch of boxes. Bees have been trained to fly through this maze and pick the correct exit. Some of these boxes, you would notice, have multiple exits. So they have to fly through each box and pick the correct exit, and finally when they go all the way to the end they get a reward.

In this particular case the bees are given a hint to navigate through the labyrinth. The hint is that they have to always pick the exit that has a green card placed beneath the correct exit hole. The way we train them is to train a bee to come to a sugar-water feeder that already has that green card pasted on it, and then we take that feeder in, one step at a time – we take it in through the entrance first and then through, say, two more steps. You don’t really have to take it through the entire maze; you take it through two or three boxes and by that time the bee has got the hang of it, that it has to actually follow the green tag that gets it to the food. And that’s what it seems to do.

Don’t worry about the details on the graph – these are just performance histograms that show you how well a trained bee performs, once it has learned the task. But they learn the task very well. Within a day they are performing this task with reliable, incredible accuracy.

[A video was shown, with the following commentary.] Here is a trained bee going through one of these mazes. This happens very fast, so you will have to watch it fairly carefully.

Here she comes, there is the green tag – and there she goes. There is the sugar-water solution in the back.

Once you have trained a bee to go through a maze like this, you can take her through any maze and she will just follow that particular rule. You can take them and make them go round and round in circles if you want to – but we would never do that!

(Click on image for a larger version)

Here is a slightly more challenging task. This is not just a dumb rule of following a tag, but to learn a symbolic signpost that will take you through the maze.

In this particular case, the bee goes into each chamber and the back wall of each chamber is decorated with a colour. That could be either blue or yellow. If the back wall is blue, it means ‘Turn right’; if the back wall is yellow, it means ‘Turn left’.

Again, bees seem to be able to learn this particular rule just as well and just as quickly. So, even though it is a more abstract rule, the performance histograms here show you that they are learning it just as well.

I don’t have a video to show you this, unfortunately, but the fact that they could do it just as well did surprise us.

(Click on image for a larger version)

Bees can learn to go through totally unmarked mazes – that is the classical labyrinth challenge – and of course the performance here is not as good as it is in the other two cases. But it is still certainly better than chance. What is interesting is that this is the only circumstance under which bees actually learn to navigate through the maze.

What I mean is this. If you take one of the bees that have been trained to navigate through the maze either by following a tag or by following a symbolic signpost, and then all of a sudden you remove the tag or the signpost, the bee is totally lost. It is only the bees that have been forced to learn this thing that actually learn the route through the maze.

In that way, bees are remarkably human. Let’s say you are driving somewhere and your passenger sitting next to you is actually navigating. You never learn the route, do you? It is only when you are forced to learn the route yourself that you pick up the route. In that way I think bees seem to behave very much the way we do.

Detecting camouflaged objects

Contrary to what some of you might think, this work is not done by a colleague who is funded by the defence services. This is just pure curiosity-driven research.

(Click on image for a larger version)

Some of you may already have seen this image, but some of you may not have. Let me ask those of you who have not seen it what sorts of objects you can see in this. If I stare at it for a while I can see the young lady, I guess I can see some birds, I can see the mountains in the background. And if I look a little more carefully I can probably see fingers, I can probably see a flute, and eventually maybe the face.

The human brain has this amazing capacity called ‘top-down’ processing, so once you have detected all the hidden objects here you will never look at this image in the same way again. Every time you look at it, these hidden objects will just leap out at you.

What is happening is that prior knowledge is there in the brain, and that causes the brain to reach out almost into the retina, pull out the object, even though it is in a very poor signal-to-noise ratio environment – the noise is very high – and extract these objects, because it knows what to look for.

We wondered whether lowly creatures like insects also possess something akin to ‘top-down’ processing.

(Click on image for a larger version)

So here is what we did. We had bees coming through this maze. They had to choose between two visual stimuli. In one case it is a textured ring presented against a similarly textured background; in the other case it is a textured disc presented against a similarly textured random background. There is a physical distance between the actual object – in one case, the ring; in the other case, the disc – and the background, so that as the bee comes in and views this object, because it sways from side to side there is relative motion between the image of the object and the background: the object is much closer to the insect than is the background.

This apparent relative motion between the object and the background allows the object to pop out and therefore be visible. Otherwise the object would be totally camouflaged against the background. You wouldn’t be able to see it.

So that’s the thing that these bees have to do: they have to disentangle this object from the background and then perceive the difference and go to the correct object.

If you train them by rewarding them always on the textured ring, for example, then try as hard as you can, train them for five days – which is a long time in the life of a bee – they simply don’t learn it. They simply cannot learn to break this camouflage.

However, if you now start afresh and train bees on uncamouflaged versions of these objects, as we have done here, using a solid black ring against a white background and a solid disc against a white background, they will learn that task quite quickly, and readily and well. And then when you subject them to the camouflaged versions of the same patterns they will immediately pick out the camouflaged objects. So clearly this prior knowledge seems to instil in them a capacity to pull out these objects from the noise, from the background.

Not only that, but once you have trained bees in this way you can give them novel camouflaged objects and they will learn to distinguish those without having to be pretrained on uncamouflaged versions. So the process of training them in this way has actually taught them that it is motion parallax that is the relative cue to pull out these objects against the background, and once they have got the hang of it, once they know what the trick is, they are able to apply it to a novel situation without having to go through the rigmarole of pretraining using uncamouflaged objects.

So bees have obviously been trained in this particular experiment to look at the world in a new way, in a way in which they wouldn’t normally look at it in a day-to-day context.

Learning concepts of ‘sameness’ and ‘difference’

Here is another aspect, which has to do with learning concepts – abstract concepts of things being the same or things being different. What I am referring to here is a thing called a delayed match-to-sample task. Human beings do that pretty well, monkeys do it pretty well, and pigeons I think do it fairly well. The idea is that you are given one stimulus, it is flashed at you on a computer screen, and after that you are given two stimuli, one of which matches what you saw before. The other is different. You are supposed to pick the one that actually matches what you saw before. So you have got to look at the original stimulus and remember what it was, and then when you are given the choice you have got to pick the one you saw before.

(Click on image for a larger version)

To see whether insects can do such a delayed match-to-sample task, we have bees coming in to a maze like the one in this diagram. We have vertical cylindrical drums, with a hole to go inside and two holes out, and a visual stimulus consisting of a yellow card to be seen on coming in. So as the bee is entering the apparatus she sees the yellow disc in there, and after she comes in she will see one yellow card and one blue card. If she picks the correct matching stimulus she gets into the chamber which contains a food reward. So the idea is that when she has seen the colour on the sample and she goes into the chamber where she sees the comparison stimuli, if she sees yellow she has to pick yellow. If she sees blue, on the other hand, she has to pick blue.

This task is something they seem to learn very well. As you can see in the performance graph here, when it is yellow they pick yellow 90 per cent of the time; when it is blue they pick blue 90 per cent of the time. So they are learning this delayed match-to-sample really nicely.

(Click on image for a larger version)

This also works with oriented line patterns, in this case gratings. They can pick the match matching the orientation and do that task equally well.

What is happening here is that the bee has to see the sample stimulus, remember it and then pick the matching one when she enters the chamber. So what happens if you increase the duration over which she has to remember this pattern? How long is the working memory operational in a situation like this?

(Click on image for a larger version)

You can test that quite readily by just extending the entrance to this configuration by having a longish entry tunnel, so we have a sample stimulus which is quite a distance from the two comparison stimuli. By simply varying the distance we can adjust the time duration over which she has to remember in order to do her task.

(Click on image for a larger version)

As you can see from this slide, when the distance is very short they can do pretty well, and as you extend this the performance drops off. The limit seems to be about five seconds, so they seem to be able to remember this particular sample stimulus for about five seconds, which is about the time you would probably be able to remember a telephone number from the time you look at it in the phone book to the time you make your phone call. And if you get interrupted during the process, you forget the number.

My colleague Shaowu is doing some really interesting experiments now where he is giving the bees interrupting stimuli, distracting stimuli, during this actual process to see how that affects their operation.

(Click on image for a larger version)

Bees can transfer this concept of matching across sensory modalities. You can train them to match scents or odours, and then give them visual stimuli, colours, and they will match the colours without having been trained to do the visual matching. So they are able to learn a concept of matching in one domain and apply it to a totally different domain.

They can also learn the concept of non-matching, which is to pick the stimulus that does not match. That is also something that is possible.

All of this is not too bad for a bee, we think.

(Click on image for a larger version)

The next thing here is associative recall. When you look at the image on this slide, do you smell lemon? I do – it works for me. Associative recall is a very powerful phenomenon. You could listen to a piece of music that you might have heard a long time ago as a kid, and that could rekindle ancient memories of what you were doing last time you heard that song. Or you could get a whiff of perfume or of cologne and it could remind you of someone you might have known a long time ago but don’t want to remember any more. Associative recall is a very powerful, deep-rooted thing that is there, and we wondered whether insects might have that too.

(Click on image for a larger version)

So here is the experiment. What we were trying to do was to see whether scents injected into the hive could trigger memories that the bee had of locations that the bee had actually visited previously and foraged at. This is work done with my colleague Judith Reinhard.

We train bees to come and visit a sugar-water feeder that is laced with the scent of rose. That is placed at a particular location, as you see on this slide. We use individually marked bees that are trained to forage back and forth between two particular sites. Then we stop that and we place a different sugar-water feeder, but now laced with the scent of lemon, in a different location and we train the same bees to forage at that feeder for a while. We know that they are the same bees because we have individually marked them, so we can be sure they are the same bees coming to this site as well.

Then what we do is to periodically alternate these two sites. When the feeder is in the first position it is scented with rose, and when the feeder is in the second position it is scented with lemon. The question is whether the bees can learn to associate each of these feeder locations with a particular scent.

(Click on image for a larger version)

Once we have done this training, which takes a couple of days, we take a break – for example, over the weekend – during which the bees can do whatever they want. They can forage wherever else they want to, because we aren’t giving them any food. But come Monday morning we set up these two dummy feeders. They are totally empty now; there is no sugar water, there is no scent. And we simply blow a scent into the hive.

In the morning there is no activity – typically it is fairly cool and there is nothing happening – but as soon as you blow the rose scent into the hive these marked bees emerge and the majority of them come to the former rose-scented feeder location. On the other hand, if you blow lemon, exactly the opposite thing happens. The same bees now emerge but they now go to the former lemon-scented feeder location.

What is nice about this experiment is that there is no scent at the actual destination. So the bees are not homing in on a scent that they sampled at the hive. What is happening in this case is that the scent that they experience in the hive is triggering navigational memories. The bee is saying, ‘Aha! Rose. I know where it is. I’ve got to fly 100 metres due north-east, and that’s where I go.’ It seems to also trigger things like the colour of the destination landmark, probably the memories of landmarks to be encountered along the way, and we think that all this is happening with the scent trigger.

And when a bee comes back home and dances, not only does the dance convey information about where the food source is but the dancing bee also passes out nectar samples to the other bees, and the taste and scent of those nectar samples will probably tell these other bees, if they have visited that particular site before, exactly where the food source is and will expedite the journey to find the food source.

(Click on image for a larger version)

(Click on image for a larger version)

This is the work of Pinar Cinar (Letzkus), a student in our lab. She has found that just as we are right-handed and right-eyed, bees are right-antennaed. So they seem to be able to learn olfactory scents and discrimination between scents much better with their right antenna than with their left antenna. It turns out that the right antenna actually carries a few more sensory receptors than the left antenna. So at least right at the front end you seem to have some biological correlate with what is happening in the rest of the pathway. That is probably not the only thing; there are probably lots of other interesting things happening at various stages of the nervous system. But right at the front end there seems to be a sensory difference between right and left.

(Click on image for a larger version)

Bees can also count landmarks on the way to a food source. I don’t have a lot of time to talk about this, but Marie Dacke, who is visiting us from Sweden, has shown that you can train bees to find food after they have encountered a particular fixed number of landmarks en route to the food source. The maximum number they can count to seems to be four.

(Click on image for a larger version)

I would like to finish by thanking my colleagues and collaborators: Zhang Shaowu, who has spearheaded a lot of the behavioural work, Judith Reinhard, who has specialised in the olfactory work, Pinar Letzkus, who has been doing the right- versus left-antenna work, Marie Dacke, who has been doing the counting experiments and other neat things, and of course Adrian Horridge, who started it all by bringing me here to Australia about 30 years ago.

Discussion

Question 1: Is the variation, or the error bars, on your measurements due to the imperfect response of an individual bee, on repeated trials, or variation within the population of bees?

Mandyam Srinivasan: That is a good point. It is a combination of both. We do find differences between individuals, and that is quite interesting. We find slow learners and poor learners as compared with good learners. We don’t discard the poor learners, of course, so their performances are all thrown into the pot. But we do find the bees can be quite individual. Even in their behaviour, even without looking at the marking on the bee – we mark them with coloured dots – just by looking at their flight patterns we can quite often tell which individual it is. They are very idiosyncratic.

Question 2: How many neurons are involved? How many neurons has a bee’s brain?

Mandyam Srinivasan: There are about three-quarters of a million neurons in a honeybee brain. I must say that a lot of these neurons are really not doing a lot of thinking work. The thinking work is probably done by about 200,000 neurons. Most of the rest of the neurons are really front-end neurons which are computing fairly low-level types of things like motion, colour and so on. But the actual cognitive aspect, if I could use that word, is done by a relatively small number of neurons, I think.

Question 3: It is so impressive, how much information and perception you have dragged out of the little bee brain, but you have probably got a lifetime’s work still ahead of you to get all the subtleties out of that system.

Do you have any plans for studying perhaps more complex brains in other insects or other animals which you can apply these principles to, with reasonable confidence that you are going to get the same kind of remarkable results?

Mandyam Srinivasan: There are several ways in which to pursue these results. I don’t know if I as such should pursue more complex brains, because a lot of supremely qualified people are doing that already. But I would like to try and push the bee brain further, if I can. One of the things I am getting very interested in is looking at the concept of emotion in simple nervous systems, and whether a simple nervous system experiences some of the basic emotions – things like joy, anticipation, disappointment, frustration, anger and so on. That’s my next goal.