Listeners:

Top listeners:

00:00 00:00

Go to album

Paul Verschure on consciousness and distributed adaptive control CSN Podcasts
Edvard Moser on grid cells and entorhinal cortex CSN Podcasts
Giacomo Rizzolatti on mirror neurons and action understanding CSN Podcasts
Robert Axelrod on game theory and prisoner's dilemma CSN Podcasts
Adrian Owen on disorders of consciousness and vegetative state CSN Podcasts
Jonathan Whitlock on markerless motion capture and posterior parietal cortex CSN Podcasts
Luis Puelles on neuroanatomy and prosomeric model CSN Podcasts
Zoltan Molnar on subplate neurons and cortical development CSN Podcasts

Alex Kacelnik on new caledonian crow and tool use

Episode 4 14.03.2013

PLAY EPISODE

Season 2013

0 Followers 10 Episodes

Description

Can a crow that has never seen a particular problem still build the right tool to solve it, and what does that tell us about the nature of animal intelligence? Alex Kacelnik explores the boundaries between insight and learning in New Caledonian crows.

Subscribe for more from the Convergent Science Network podcast series.

Alex Kacelnik brings a biologist’s perspective to animal cognition, positioning intelligence as an evolved toolkit shaped by natural selection rather than an abstract capacity to be ranked on a human-centric scale. He draws a critical distinction between risk, where probabilities are known, and uncertainty, where even the nature of the problem is unclear, arguing that learning transforms individual uncertainty into manageable risk by filling in the parameters that evolution could not anticipate.

The centerpiece of the discussion is the New Caledonian crow, the most intensely tool-dependent non-human species known. These birds manufacture at least five categories of tools including hooks, straight sticks, and elaborately shaped pandanus leaf strips, showing regional variation that suggests cultural transmission. In laboratory settings, the crows demonstrate remarkable flexibility: they select tools of appropriate length and diameter for specific problems, build hooks when straight tools will not work, and solve multi-step problems requiring sequential tool use on a trial-unique basis. Kacelnik emphasizes that these behaviors cannot be fully explained by chaining previously reinforced responses, as the complete sequences have never been experienced before.

Yet Kacelnik resists easy mentalistic interpretations. He positions himself closer to the “killjoy behaviorist” than the “mystical psychologist,” insisting that terms like insight, planning, and understanding should only be used when backed by algorithmic models of how experience translates into novel solutions. A key experiment illustrates this principled caution: crows could innovate by dropping stones into a mechanism to release food, but only if they had prior experience with how the magnetic release mechanism worked. Innovation requires partial knowledge as scaffolding, not magical leaps of comprehension.

The episode also examines how crows use tools not just for food extraction but for exploring potentially dangerous objects at a safe distance, and how sexual selection in siskins illustrates the complex evolutionary pressures shaping cognitive abilities across bird species.

Tagged as:
Crows Hooks Straight new caledonian crow tool use

About the author

CSN Podcasts

Both the triumphs of humanity and its most evil deeds have resulted from collaboration. In a time where humanity is required to aspire to the former and minimize the latter, the question arises of how collaboration arises and why it fails. Surprisingly, this phenomenon, so central to who we are, is not well understood. Hence, a collaborative effort is required to understand collaboration in its full biological, psychological, sociological, cultural, and economic complexity and to translate this understanding into operational impact. This series of podcasts is one step toward achieving these complementary goals. The Collaboration Podcast presents interviews with people who are central orchestrators of collaboration in various domains including business, government, science, art, health, sustainability, and the military. The discussions were conducted by Prof. Dr. Paul F.M.J. Verschure and members of the Program Advisory Committee of the Ernst Strungmann Forum on Collaboration (https://www.esforum.de/forums/ESF32_Collaboration.html) during 2021 and had the goal to sketch a map of opportunities, challenges, and obstacles in human collaboration. The forum took place in May 2022, and now we would like to share this series of interviews with a broader audience. The full report of the Forum will be published in 2023 by MIT Press. The podcast was produced by the Convergent Science Network (https://www.convergentsciencenetwork.org/). Context: The stability of social systems depends critically on realizing sustainable methods of “collaboration,” yet how and by which means collaboration is achieved is not clearly understood; neither are the conditions or processes that lead to its breakdown or failure. Collaboration can be understood as cooperation between agents toward mutually constructed goals. Part of the reason for our lack of understanding is that the phenomenon of collaboration is, by nature, a highly multidisciplinary problem, and effective research into its complexities has been difficult to achieve across the broad range of scientific and technical disciplines involved. The need for a fundamental understanding of collaboration, however, has become increasingly important. Not only does humankind demand answers as it attempts to address critical challenges at multiple scales (e.g., climate change, migration, enhanced automation, social and economic inequality), but ever-increasing technological and economic means of interconnecting people and societies are disrupting long-established, familiar patterns of how we interact. Radical technological changes that are ongoing have the potential to reshape collaboration in ways that are currently hard to predict or influence (e.g., by altering configurations in interaction, information creation, and modes of communication). On one hand, such changes could disrupt hitherto stable forms of collaboration by affecting critical communication channels and traditional roles, as can be observed in the rapidly changing patterns in governance, commerce, and social interaction. Conversely, technology could lead to the emergence of novel, successful forms of collaboration that deviate from traditional “hierarchical” architectures. Evidence of this can be seen in areas as diverse as highly automated manufacturing plants, the open science movement, collaborative software repositories, user-centered services, and the sharing of economy-based modes of organization. Without a fundamental understanding of the mechanisms, processes, and boundary conditions of collaboration, it is not possible to evaluate or predict which of these possible scenarios are sustainable or even plausible. The Forum “How Collaboration Arises and Why it Fails” (May 8–13, 2022, Location: Frankfurt am Main, Germany) Chairs: Andreas Roepstorff and Paul Verschure Program Advisory Committee: Jenna Bednar, Julia R. Lupp, Bhavani R. Rao , Andreas Roepstorff, Ferdinand von Siemens, and Paul Verschure

Timestamp

00:00:03 - This is the Convergent Science Network podcast. Leading researchers in the domain
00:00:10 - of neuroscience, brain theory and technology are interviewed by Paul Verschoor and Tony Prescott.
00:00:25 - This is Paul Verschoor with the Convergent Science Network, work together with
00:00:29 - Tony Prescott and our guest Alex Kacelnik.
00:00:33 - And Alex is a biologist who has been studying in great detail the extraordinary,
00:00:39 - also cognitive capabilities of a range of animals, in particular birds.
00:00:44 - So Alex, why do you think birds are a good target species to study to understand animal cognition?
00:00:51 - Well, as in many of these things, some of these things have history more than
00:00:56 - real logical explanation. I mean, one starts in a system and keeps going.
00:01:01 - But in reality, birds have certain good properties from the point of view of studying behavior,
00:01:09 - in the style that I'm interested in, because A, they are diurnal,
00:01:15 - and so you can study them in daytime, and they are very visible.
00:01:19 - So if you want to study them in the field, you can see what they are doing,
00:01:23 - as opposed opposed to seeing rats, which you have to go down into the sewage
00:01:26 - and studying them at dark.
00:01:30 - So it's much more complex.
00:01:33 - So birds have a wealth of stereotype behavior, which has occasionally been misinterpreted,
00:01:41 - very often been misinterpreted as implying lack of the capability for flexible behavior.
00:01:47 - And when, in fact, having the ability to do certain things without having learned
00:01:51 - them doesn't mean that you cannot combine them with innovation and new discoveries
00:01:59 - and cognitively demanding tasks.
00:02:02 - So I started, I guess, well, if I ignore my very early years,
00:02:11 - I started working on birds because I could...
00:02:17 - Look at them both in the field and in the lab, I could create conditions for
00:02:23 - relatively simple interfaces.
00:02:26 - They don't use, although they
00:02:28 - use more than ghosts believe, they don't use smell as much as mammals.
00:02:31 - It doesn't mean that they don't use olfactory cues, but they are less olfactory
00:02:35 - than mammals, meaning that visual stimuli, which we can easily understand and
00:02:40 - program, can kind of easily be implemented in the laboratory.
00:02:47 - And they have a life cycle, regular life cycle in a way.
00:02:54 - But now you also placed that study of cognition in a very clear evolutionary perspective, right?
00:03:01 - So you really see that cognition evolved to actually address a very specific
00:03:06 - set of features of the interaction with the world, which is that the kinds of
00:03:11 - problems we encounter to survive might come in different qualities.
00:03:14 - So, what's the distinction that you see there?
00:03:17 - Yeah, you're absolutely right. I see myself fundamentally as a biologist,
00:03:21 - and I see cognition as yet another part of the toolkit of the animal to adjust to its environment.
00:03:29 - In a sense, it's not different from the way Herbert Simon conceived the understanding of the mind.
00:03:36 - I mean, you have, you know, the famous metaphor of the scissors where you have
00:03:42 - the two blades to make the scissor cut.
00:03:45 - And then if you think of understanding the mind without understanding the environment
00:03:50 - in which it evolves, not knowing which problems the mind has evolved to solve,
00:03:54 - it's going to be a handicap for you to understand the mind.
00:03:57 - The mind, and I use the term very loosely here, but they say the product of cognition in general,
00:04:05 - has not evolved to do the fanciest possible calculations or to just give joy
00:04:14 - to either observers or the owners of those minds.
00:04:20 - It has evolved because it's a good way to solve problems, and the problems are
00:04:24 - determined by natural selection. So you try to link both and evolve towards
00:04:31 - a better understanding.
00:04:33 - Right. But it calls me the case that in evolution, let's say.
00:04:37 - You can get away with pre-wiring stereotype behaviors for problems that at an
00:04:42 - evolutionary time scale are constant.
00:04:45 - Let's say to right yourself against gravity might be something you could just
00:04:50 - pre-wire because gravity doesn't change so rapidly,
00:04:52 - while finding the fridge to get a beer might be a problem for which you might
00:04:57 - want to have at least to some extent cognitive capabilities, right?
00:05:01 - Yes, you're absolutely right. In the In the area of decision-making,
00:05:04 - there is a very crucial distinction between what is normally understood as risk
00:05:10 - and what is understood as uncertainty or 90 and uncertainty.
00:05:14 - In the case of risk, you know the probabilities of events.
00:05:17 - And although the world is inherently stochastic, that is, you cannot predict
00:05:21 - exactly the consequences of your actions, you know the probability distributions around yourself.
00:05:27 - In the case of uncertainty, you may not even know the nature of the problem
00:05:31 - very well. You don't know the probabilities of each part of the problem.
00:05:36 - And in evolutionary sense, things that for the individual are a case of uncertainty,
00:05:42 - for evolution may be a case of risk.
00:05:45 - And that in one lifespan, you may not know the probability that,
00:05:49 - say, it's going to be warm in May if you are born in February, for example.
00:05:55 - But evolutionarily speaking, you can know that because that's encoded in your
00:05:59 - genome gnome through the success of your ancestry.
00:06:03 - But also the first model animal that you...
00:06:06 - Presented to us was the siskins, right?
00:06:10 - So you were looking at this problem of mate selection in siskins and you saw
00:06:17 - this as an example of also how ambiguous it is to think about,
00:06:23 - cognition as necessarily helping you in survival, right?
00:06:27 - So what was the message of these experiments with these siskins?
00:06:30 - Well, I was referring to experiments done by the group of Senar,
00:06:35 - Juan Carlos Senar, here in Barcelona.
00:06:36 - And what they did was to study these little birds in which some aspects of female
00:06:43 - mate choice are very well understood.
00:06:45 - The females prefer males which have an enhanced yellow bar in their wings.
00:06:52 - And what they did was to ask the question, why do the females prefer this?
00:06:56 - And they tested the animals in the laboratory and they found that And there's
00:07:00 - a correlation between their ability to solve certain problems,
00:07:03 - how fast they solve it, and the feature that the females are using to select between them.
00:07:10 - So as if females are using the yellow bar as a proxy for picking up a male that's
00:07:16 - going to be good at feeding their young, for example, or at providing their
00:07:20 - young with good genes for being efficient foragers.
00:07:24 - I simply was using their work as an example of the complexities of saying that the male.
00:07:32 - Intelligence may evolve through sexual selection the female role in this but
00:07:37 - the females is actually using a characteristic that we don't fully understand to,
00:07:45 - Develop their preferences, right? So but then I.
00:07:51 - On the other hand, that raises the question, why are not all these syskins,
00:07:59 - let's say, developing a big yellow bar even though they're stupid?
00:08:05 - Because you would easily try to, in that sense, fool. You could fool your potential
00:08:09 - mates. Why is that not happening with these syskins?
00:08:12 - Well, that's taking us, your question takes us to a basic issue of the evolution of communication.
00:08:16 - In communication, you have an emitter and a receiver. Some signal has evolved
00:08:21 - in the emitter because it's a good way to alter the behavior of the receiver
00:08:27 - through the receiver's senses.
00:08:30 - And the question is, it's no question I would do whatever favors me.
00:08:35 - But, of course, the receiver evolves responding to signals in a way that actually benefits it.
00:08:43 - And so why would a receiver accept the signal if there's a frequency-dependent
00:08:48 - issue immediately emerging?
00:08:49 - If too many stupid males had the sign for intelligence, then the preference
00:08:59 - for that clue would disappear and the females would be selected out.
00:09:03 - It wouldn't convey information.
00:09:04 - So you have a constant balance between the advantage of the psychology of the
00:09:12 - receiver and the benefit to the emitter.
00:09:14 - So communication is not about accuracy of conveying information.
00:09:19 - It's about efficiency of manipulation.
00:09:21 - And for the receiver, it's efficiency of decoding variables in the emitter that are of use to self.
00:09:30 - And I was just trying to illustrate a little bit of that. Right, exactly.
00:09:35 - So then the core of your experiments that we're going to talk about later with
00:09:40 - a number of birds, bird species…,
00:09:46 - It all turns around this issue of problem-solving and insight.
00:09:49 - And you started with this very classic example of colors, chimpanzee experiments.
00:09:55 - So why is that such a critical experiment or such a strong illustration of the
00:10:01 - kind of paradigm that you're interested in?
00:10:04 - Just to remember, these experiments were done in the early 20th century.
00:10:10 - And the basis was to present the chimps with a novel problem and see how they
00:10:16 - innovated with instruments around themselves to achieve a solution.
00:10:21 - At the time, that was interpreted as evidence for the existence of insight and
00:10:27 - of mental modelling and trial and error in the mind by the chimps.
00:10:33 - And I use this as an example of the weakness of our tendency to project the
00:10:40 - way we think we solve problems into the data we collect from other species.
00:10:46 - I use as an example of this some work from some of the people of the American Behaviorist School,
00:10:54 - the Scandinavian School, particularly Robert Epstein, who showed that if you
00:10:58 - train pigeons on the components of a task, and one day you offer a problem that they have never faced,
00:11:04 - but where the different components can be chained together to achieve a solution, the pigeons also did it.
00:11:11 - And of course, this had not been done systematically in the chimps' experiments,
00:11:15 - but the chimps have their personal experience, which they could do it.
00:11:19 - And the issue I raised was, if the pigeon does the same as the chimp,
00:11:24 - should we conclude that none of them is intelligent or that both of them are?
00:11:29 - And how do we separate them? And that was just a springboard to start designing
00:11:34 - cases where we see animals solving problems,
00:11:39 - which are novel in the way they are presented, but somehow resonate with their
00:11:43 - experience, obviously.
00:11:45 - And how could then unravel, really, what is the task for the organism?
00:11:50 - How is it doing? And this is where we connect with people working with relatively autonomous robots.
00:11:56 - But now, how I take your Epstein example is that,
00:12:02 - so Epstein had his four equations of what he called generative behavior,
00:12:06 - which are very reminiscent, obviously, of also Thorndike's law of effect.
00:12:11 - And that essentially means that as a learning organism, you're really at the
00:12:15 - mercy of your environment.
00:12:16 - Your environment is instructing you about what you have to do.
00:12:19 - And to build a chain, you do have to receive specific reinforcement that tells
00:12:24 - you, well, A and B belong together and should be executed together. Right.
00:12:28 - So that's sort of this empty organism notion.
00:12:34 - But now, in your case, if you look at insight, insight might require something
00:12:39 - like an internal model, internal representation.
00:12:42 - So where do you place your own thinking in between those extremes?
00:12:46 - Do you see this as relying on internal models, or do you really think that this
00:12:51 - kind of complex behavior,
00:12:53 - problem-solving behavior that might look like insight can really be instructed
00:12:58 - by direct reinforcements from the environment as the behaviorists would have it.
00:13:02 - Well, there are several aspects to your very complex question.
00:13:07 - One of them is that a totally instructed organism doesn't really exist.
00:13:12 - I mean, as you well know, we filter the information that the world is giving us.
00:13:17 - We selectively take particular relations in the world.
00:13:21 - I mean, we have a selective perception, selective processes of conceptualization.
00:13:24 - And different species differ in what they filter of their environment and how
00:13:29 - different in their gaze and then how they look at the world.
00:13:32 - And as a biologist, I'm very keen on taking that into account.
00:13:37 - What is it that the organism is actually trying to extract as significant information?
00:13:41 - On the other hand, the job that learning is doing for the animal,
00:13:45 - if we go back to the notions I mentioned before of risk and uncertainty.
00:13:49 - What learning is doing is transforming uncertainty into risk.
00:13:52 - So you start an animal with not knowing the problem it's in,
00:13:58 - other than with what natural selection has actually encoded it for,
00:14:02 - and evolving by its own experience into actually plugging in the parameters
00:14:09 - of its real life circumstances.
00:14:12 - I mean, what world I am in, and what are the affordances of the objects around
00:14:19 - me, and what are the laws of this, even the social environment in which I'm moving.
00:14:24 - And that cannot be anticipated by natural selection, so that job is done by learning.
00:14:29 - Now, you asked me about insight. Right.
00:14:33 - I don't want to really define myself as, it's not the kind of thing you can be for or against.
00:14:40 - I mean, in a lot of cases, it is very frequent to over-interpret data as demonstrating
00:14:46 - that this could not be done other than by insight.
00:14:50 - But in many cases, insight is just labeling something like the animal could
00:14:56 - not do it, and then it did it.
00:14:57 - And this discontinuity in
00:15:01 - the data is explained by something miraculous that happened
00:15:05 - in the animal that we don't understand better by putting a name to it so if
00:15:08 - we could actually show what mental operations the animal is doing and have some
00:15:14 - handle experimentally or quantitatively in some kind of way then i i'm happy
00:15:19 - with mentalistic interpretations but they have to have a handle Right, exactly.
00:15:53 - Of an image then it takes longer to actually
00:15:56 - do it the greater the angle and if you ask individuals why is how they do it
00:16:04 - they tell you that in their minds they are rotating the the object until they
00:16:09 - see it vertically and then they can actually take a decision so So,
00:16:13 - Delius did the same experiment,
00:16:18 - but found that the pigeons had no extra time for objects that were rotated with
00:16:25 - respect to the training orientation.
00:16:30 - And if you did find it, you could conclude and accept that maybe they were using
00:16:35 - some kind of mental rotation and do further experiments about it.
00:16:39 - But the fact that you didn't,
00:16:41 - doesn't tell you that they don't have the capability for mental rotation.
00:16:45 - It tells you that an animal that flies looking at a horizontal world from above
00:16:51 - and looking from a different perspective doesn't have a priority axis like the
00:16:56 - vertical one and can actually quickly identify images by their pattern regardless
00:17:01 - of whether they are rotated.
00:17:03 - But we can't actually tell what's going on in the mind of the animal.
00:17:08 - But is your statement there also more that you're saying, look,
00:17:11 - you want to link to Epstein and this whole tradition behind him,
00:17:14 - going back to Thorndike and Pavlov, not so much to say, look,
00:17:18 - conceptually, I believe in an empty organism, but in our explanations,
00:17:22 - we should keep it as minimal as possible and really ground it in the data we
00:17:26 - have and not in our anthropomorphic interpretations of this behavior of the
00:17:29 - animals. Yes, absolutely.
00:17:32 - I think from that point of view, I would place myself in the range between the
00:17:37 - mystical psychologist and the killjoy behaviorist.
00:17:44 - I would place myself closer to the killjoy than to the mystical.
00:17:50 - But i admit that a
00:17:53 - lot of a behavior we ourselves observes and collect
00:17:56 - on our animals we don't have an algorithmic model to
00:18:00 - say how does experience translate into these problems also
00:18:03 - and we are looking at these problems to see if we can
00:18:05 - build them yes that was more or less my question but you're saying that the
00:18:10 - key principle here is is parsimony so we look for the simplest explanation of
00:18:14 - behavior and a second principle i guess is continuity that we look for things
00:18:19 - that we see in other animals that might explain this when we see it in birds.
00:18:25 - Is there not some reason to expect maybe that some of these bird species do
00:18:31 - have competences that might not exist in many other animals?
00:18:34 - Should we not be a bit more generous in expecting birds, especially ones with
00:18:40 - larger brains, to have some of the abilities that you might only see, say, in primates?
00:18:46 - Yes, I think that's what you say is empirically validated already.
00:18:52 - I mean, we know already that many problems that others and ourselves have shown
00:19:02 - birds to be capable of solving have not yet been shown as being solved by mammalian
00:19:09 - species, including primates.
00:19:11 - There's no question that every species has particular competencies.
00:19:17 - And if we create a scale where the yardstick is human ways of doing things,
00:19:24 - then the closer you are to humans, the better you're going to score in that
00:19:28 - scale. But that's circular.
00:19:31 - But their target animal in many of your experiments has been the Caledonian crow.
00:19:38 - And what makes that animal species so special?
00:19:42 - Well, the New Caledonian crow is, and I'm now including the great apes,
00:19:48 - is the most intensely dependent tool user among non-humans that we know of.
00:19:58 - They show a species-wide tendency to use tools. They use it in nature.
00:20:05 - They fulfill an important part of their ecology. technology and
00:20:08 - they don't use it in a rigid fashion
00:20:11 - we know that it's not just that they have inherited
00:20:14 - a set of motor patterns they have an
00:20:17 - inherited predisposition to apply
00:20:21 - tools to problems they face and in some respects as juveniles they show a kind
00:20:29 - of proto movements of what is going to be the adult tool use and they do it
00:20:35 - for something that we might call play if we saw it in a young human.
00:20:40 - But at the same time, they have the capability for cultural transmission so
00:20:44 - that we know that they can learn from others to some extent.
00:20:50 - They have regional variation in the tools that they are using,
00:20:55 - and that could be connected to a physical culture, although that hasn't been demonstrated yet.
00:21:00 - So all these things make them very interesting. But a typical example would
00:21:07 - be that their friends would take a stick.
00:21:09 - To fish, as you described yourself, for larvae in tree, in the mask of trees, right?
00:21:16 - This would be the typical thing, how they would use a tool.
00:21:19 - Yes. Actually, local people, non-scientists, knew this for a long time.
00:21:24 - And Gavin Hunt, working from New Zealand, brought it to the attention of science
00:21:30 - in the late 90s and published some very interesting papers actually describing what they do in nature.
00:21:39 - And they use different kinds of tools, at least three or four different kinds of tools,
00:21:47 - five maybe, depending on how you categorize a kind of tool, like hooks or straight
00:21:52 - sticks or blades of pandanus plants,
00:21:56 - that in all cases, they build themselves and they modify and then they use for extracting food.
00:22:06 - We don't know to what extent in nature they are selective of producing tools
00:22:11 - that are appropriate for a particular problem that they are facing.
00:22:17 - But we know that by taking them to the laboratory and giving them different
00:22:20 - problems, they can do this.
00:22:22 - They can build tools which are appropriate for the problem they have in hand.
00:22:27 - But now if you say that they have about five tools they would use, this is like a stick?
00:22:32 - I said five categories of tools. Yes, because they're very different, all of them.
00:22:36 - But for example, they can use segments of a flexible vine that they cut with
00:22:44 - the appropriate length and poke it into holes with the teeth facing backwards
00:22:50 - so that they can actually rake things out.
00:22:52 - They can produce hooks by
00:22:56 - sculpturing twigs at the branching point
00:22:59 - and so leaving just a short segment
00:23:02 - on one side and a longer one in the other and we
00:23:06 - know in the laboratory but not in nature how
00:23:08 - these hooks are actually used they build
00:23:12 - the most complex tool of
00:23:17 - any animal really
00:23:21 - alien invertebrate if you exclude things like spider webs
00:23:24 - by actually cutting the
00:23:28 - edge of blades of leaves
00:23:31 - of pandanus trees which are like these leaves are like flat surfaces and they
00:23:38 - cut the edge with a particular step shape format so that they are thicker more
00:23:44 - robust on one side and they taper to the other and they use this for.
00:23:50 - Extraction purposes but we don't know for example why in
00:23:53 - some areas they do it and in others they don't whether it
00:23:57 - is and also why their shape is different whether
00:24:00 - it's functionally different or it's just cultural drift that leads honey that's
00:24:06 - what i was word so so interested in so if we have these different categories
00:24:08 - of tools now if you talk about cultural variation what should i think of given
00:24:14 - these categories of tools what's the culture variation here cultural variation is for For example,
00:24:19 - this is again work by our colleagues in New Zealand, that they found that the
00:24:25 - typical pandanus leaves that they produce,
00:24:28 - in some cases have a perfectly rectangular shape, which is kind of broad.
00:24:35 - In others, they have a thin one. And yet in other places, they have a stepwise
00:24:41 - one where they start thick and they end fine, which is a more advanced, more useful tool.
00:24:47 - In some areas, the three kinds of tools are built, which is,
00:24:51 - in a sense, a contradiction with the notion that there is a better,
00:24:55 - there's a best kind of doing it.
00:24:56 - But definitely what is the case is that the predominant kind of pandanus leaf
00:25:04 - tool is different in different regions.
00:25:08 - Now geographic variation does not prove culture, but it's an obvious associated concept.
00:25:16 - I was wondering, do the birds show anticipation?
00:25:20 - So if they're going to go on a particular kind of foraging trip,
00:25:23 - will they go and make a tool and take it with them?
00:25:25 - And also, do they store tools and reuse them?
00:25:29 - Well, we have shown through the use of cameras.
00:25:35 - With some of my colleagues, particularly Christian Roots, who's now in Scotland.
00:25:39 - Used to be in Oxford at the time, We managed to put cameras on birds which are
00:25:46 - moving freely in the woods and could see some of that.
00:25:50 - For example, we could see that they could make a tool and take it hundreds of
00:25:56 - meters away somewhere else and use it.
00:25:59 - But that doesn't give us any evidence that the animal knows what kind of problem
00:26:05 - it's going to face when it arrives and is doing the right kind of tool here.
00:26:09 - But in the laboratory, we know that when they face a problem,
00:26:13 - for example, where they need a hook to extract food, they actually make a hook
00:26:18 - when some other shape of tool would not work.
00:26:21 - So they really seem to be able to do this kind of planning.
00:26:24 - Similarly, when we give them a task in which they need to collect one kind of tool,
00:26:30 - say of a given length, to pick up another one that can be used to pick up yet
00:26:35 - another one, that kind of complication, then they are capable of doing it in the lab.
00:26:39 - And if you think of how you would program an autonomous machine to do this,
00:26:45 - you couldn't do it without some kind of building enrichment of the reinforcement experience.
00:26:53 - Just purely repeating what works doesn't take you to the end because the animal
00:26:58 - has never, these are trial unique things where they've never done these complete
00:27:02 - sequences. So they have to do something equivalent to what one would call planning.
00:27:06 - But I resist as much as I can to use words which are heavily laden with meaning,
00:27:14 - like planning or understanding or insight,
00:27:18 - when all they are doing is calming our anxiety about not knowing what the animal
00:27:22 - is doing by putting it a label.
00:27:24 - So, yes, they do some planning in terms of the behavior they do anticipates
00:27:28 - the problem rather than acting by consequences.
00:27:32 - Before we sort of get into the planning bit,
00:27:35 - I mean, you have gone through quite a series of experiments to really document
00:27:41 - in detail the tool use that these crows are capable of and how they can generalize.
00:27:47 - So what are the key features of their tool use that stand out in your mind?
00:27:54 - Well, the first one, and I was anticipating a second ago, is that not only...
00:28:00 - Everything they do is contained, in any obvious way, in the experience they had before.
00:28:06 - It does appear as if we need some kind of model by which the animal does what,
00:28:11 - let's say, the person in the street would call understanding of the problem,
00:28:15 - and I don't have a better label for it, but at the same time,
00:28:19 - I'm aware of the difficulties of using such term.
00:28:22 - The animal sees a problem and produces a solution that it has never experienced before.
00:28:28 - It's not just repeating with greater frequency what has worked.
00:28:32 - But on a trial-unique basis, it's solving problems one after another.
00:28:36 - That's a common feature. At the same time, we find that they need knowledge,
00:28:42 - which in many cases is completely logical to solve it.
00:28:45 - I can give you one study, for example. It
00:28:49 - was shown in other corvids by colleagues in Cambridge that rooks are capable
00:28:57 - of discovering how to drop stones in an instrument to dislodge a magnetically
00:29:04 - held platform to release a reward. world.
00:29:07 - So that was extraordinary in itself. But the Cambridge rogues had done this
00:29:14 - task by being first trained to drop stones accidentally.
00:29:19 - They could see that and then they could innovate by picking up stones from a
00:29:24 - distance and bringing it there.
00:29:26 - So what we wonder is how could these These animals know that a stone would actually
00:29:34 - dislodge the magnet without having experience of how the apparatus worked.
00:29:39 - So what we did was to...
00:29:42 - Give different groups of animals either experience with dislodging the magnet
00:29:48 - by pecking directly at it or not.
00:29:51 - And we placed them without any previous experience of stones dropping in front of the machine,
00:29:57 - and the ones that knew how the magnetic box operated could solve the problem,
00:30:03 - could innovate by bringing the keys like the rooks had done.
00:30:07 - But the ones that didn't know that contingency just looked at it and couldn't
00:30:11 - do it. So you can innovate, but you have to build on partial knowledge.
00:30:17 - Would you call it scaffolding? I would call it that, but self-scaffolding in
00:30:23 - the sense that the animal is constructing on the basis of partial elements of behavior,
00:30:28 - which is perhaps not surprising because animals don't have any reason to understand
00:30:32 - a machine without having some possibility of interacting with it,
00:30:37 - and in this case, a delivery box.
00:30:39 - But now, don't we also face a challenge here because we look at the task and
00:30:43 - it looks really complex.
00:30:44 - But in some sense, there's also an invariant in all these tasks because it's
00:30:48 - always retrieving a food item from, let's say, a tube-like structure, right?
00:30:55 - Which is in some sense a condition for which these animals might have been optimized
00:30:58 - because that's how they have to eat, find their food in the trees where they live.
00:31:02 - So maybe to us it looks very complex. but maybe for these crows it all looks
00:31:07 - like the same problem that they always solve in the same way which is get a stick, get a stick like,
00:31:12 - get a stick-like object, and start poking in that hole.
00:31:16 - No, this is not an accurate description of what's going on.
00:31:23 - I have different lines of argument here. One is that not all our experiments
00:31:28 - are about extracting food.
00:31:30 - In some of them, we give the animals unknown, potentially threatening objects,
00:31:36 - and rather than touching them immediately with their beaks, they pick up a stick
00:31:42 - and touch them at a distance.
00:31:43 - So they use tools to acquire knowledge about the world in which they are,
00:31:48 - in addition to using them to extract food. That's one line of argument.
00:31:53 - Another is that the tasks actually are extremely different in that some of them
00:31:58 - require selection of the right tool that would go through a hole, for example.
00:32:03 - We give them food in a tube which which has different holes and natural twigs
00:32:10 - that have to be sculptured to the right diameter in order to pass through the hole.
00:32:16 - They can do that. Or they have to choose the right length. Or they have to push as opposed to pull.
00:32:22 - And they learn these kind of things.
00:32:25 - There are many, many different topographies of the problem that they face.
00:32:31 - Alex, I could still argue for the sake of it that each of these examples you
00:32:37 - give me, I could decompose in terms of a stereotype behavioral pattern that
00:32:41 - is modulated in some way, right?
00:32:43 - I could say, well, also during evolution, they have learned to not approach
00:32:47 - snakes too quickly, so they always use a stick to do that.
00:32:50 - But that's a very discrete, well-defined situation where they do it.
00:32:54 - So what is the common feature of all these tasks that makes you believe that
00:32:59 - you really have to think about a fairly rich internal model and insight, right?
00:33:05 - As opposed to, let's say, also if you want a more behaviorist decomposition
00:33:10 - in more stereotyped reactive behaviors.
00:33:14 - Well, I'm hesitant about this question you're asking because on one hand,
00:33:19 - you told me that all these tasks are very similar.
00:33:24 - And then you say that they all decompose onto many different elements.
00:33:29 - I wonder what kind of human behavior cannot be decomposed also in that particular way.
00:33:34 - It's a typical behaviorist challenge, right? Yeah. Sure. Yeah.
00:33:37 - And so I think I would start by saying that the very diversity of problems that
00:33:43 - they can solve and the fact that they can solve trial-unique problems with different elements.
00:33:49 - Components, is an indication that the model we need is richer.
00:33:56 - I'm still completely in agreement with you, and I said that from the beginning
00:34:00 - today, that I'm not after finding a nucleus that we are going to call mind and
00:34:07 - we're going to say, we can't go no further.
00:34:09 - This is it. We come to this point and from then on the animal thinks.
00:34:14 - So that's not what I hope to do.
00:34:17 - What What I hope to do is to find a sufficiently rich description of the behavior
00:34:23 - of the animals and the problems they solve, that we do it algorithmically.
00:34:28 - I don't think that the basic principles of combining associative learning,
00:34:33 - all forms, you know, instrumental and Pavlovian conditioning,
00:34:38 - are going to be sufficient for this.
00:34:41 - But we know that already. We know that rats learn about the temporal location
00:34:45 - of stimuli even if they precede contingencies.
00:34:49 - We know that the pigeons can form concepts of very high level.
00:34:55 - They can decode, for example, some of the Japanese work,
00:35:02 - Watanabe and others, that show that they can distinguish a cubist from an impressionist
00:35:07 - painting, painting, even with paintings that have never seen before,
00:35:11 - simply by being trained with those categories of stimuli.
00:35:15 - Or we can think of the classic Tolman experiments around the cognitive map,
00:35:19 - showing latent learning of very complex internal representations.
00:35:22 - Exactly, absolutely, yes. As you know, some of the people who have developed.
00:35:29 - More sophistication in reinforcement learning, like Satan and Bartholomew and
00:35:32 - that kind of people, have actually shown that some of the Ptolemian type of
00:35:37 - learning could be reproduced by some enhancement to reinforcement learning.
00:35:44 - But my personal experience is, yes, some can, but some is not quite plausible
00:35:51 - in terms of being the way the animal does it. But I don't have a better alternative.
00:35:55 - Okay. I think to some extent in psychology and animal behavior,
00:36:02 - there is a little bit of an assumption that a relatively easy thing to do is
00:36:08 - association learning and a relatively hard thing to do is planning.
00:36:12 - And that kind of came out of behaviorism. And then the criticism was that,
00:36:16 - well, all these great human behaviors that we do can't be explained by that.
00:36:20 - But then in AI, people discovered very quickly that actually some aspects of
00:36:24 - planning are relatively easy to do.
00:36:27 - So things like means-end analysis was an early discovery by Herb Simon.
00:36:32 - You could program this in a computer, as long as you had the right decomposition
00:36:36 - of the problem space, which we can come back to. too.
00:36:39 - But I mean, so coming at this from the point of view of robotics,
00:36:43 - it's sometimes a little bit hard for me to understand why biologists are so
00:36:48 - reluctant to say, well, means-end analysis is something that an animal brain might be doing.
00:36:54 - Because from the point of view of the computational problem,
00:36:59 - it's perhaps not such a difficult challenge as some of the other things that
00:37:03 - our brains do really well.
00:37:04 - Yes, I agree with you. But I think that when you say biologists,
00:37:09 - I think that some biologists are excessively tolerant of such explanations,
00:37:15 - and other biologists are excessively reluctant to use them.
00:37:20 - And I'm not saying that I'm in the just middle, because I couldn't possibly
00:37:27 - say that. But the point is to add.
00:37:30 - Each process to the extent that you can test it and test its properties experimentally
00:37:37 - and model it and that kind of thing.
00:37:39 - So associative learning for me is a good candidate on the first line of battle,
00:37:46 - not because it's easier necessarily.
00:37:49 - Because in many cases we don't understand how some of this happened,
00:37:53 - but because we know a lot about it and we know that it explains a lot of the data satisfactorily.
00:37:58 - You can get the right parameters, you explain many different kinds of things,
00:38:02 - while planning or
00:38:06 - insight are descriptive processes that
00:38:10 - don't by themselves link to anything that you
00:38:13 - can actually immediately quantify and simulate
00:38:17 - as a process or write in the form of code and because
00:38:21 - of that in as much as a hypothesis has
00:38:25 - a reassuring ring to it i don't
00:38:29 - favor it as a first thing i would like i prefer
00:38:32 - hypotheses that actually stick their neck out and say well if it is this then
00:38:38 - this is what you should be seeing and so if you look at some of the early ai
00:38:43 - which was doing this sort of search space techniques and uh means sense analysis
00:38:48 - and people People like Herb Simon were saying,
00:38:49 - well, look, let's see how people, not Herb Simon,
00:38:53 - but experimentalists saying, let's look at how people solve,
00:38:57 - say, the missionary and cannibals problem,
00:38:58 - which is a means-ends analysis type task where you have to solve a difficult
00:39:06 - leap, which involves not taking the obvious next step.
00:39:09 - And so there are experimental techniques there for trying to identify when people,
00:39:15 - humans, are using these kind of planning techniques.
00:39:19 - So I'm just trying to wonder whether there is maybe more a positive agenda.
00:39:26 - That we could have for you know let's assume more of about bird cognition and
00:39:32 - see if we can design experiments that that really push uh push the possibilities
00:39:38 - of what they're doing in terms of.
00:39:42 - Identifying how there are similarities between what they're doing and maybe
00:39:45 - what humans are doing an example might be uh analogical reasoning you know can
00:39:50 - you can you develop a way of designing, maybe this has been done already,
00:39:56 - you design one apparatus which involves a sequence of moves,
00:40:01 - that the bird might discover,
00:40:05 - and then you design something which is visually very different,
00:40:09 - but there's some deep homology between the sequence of actions that you did
00:40:12 - over here that you can use over here.
00:40:14 - I mean, has that been done, or is it realistic to think about trying to do that?
00:40:18 - Well, it has been claimed to have been done in some cases.
00:40:24 - I am completely sympathetic to your view that we should have a positive agenda
00:40:31 - in that respect and trying to see, well, if the animals are using what in AI we call planning, say,
00:40:38 - then we should see these properties or we should also be capable of doing these
00:40:42 - tasks that we wouldn't have thought otherwise.
00:40:43 - So, it's not that this hypothesis or even things like, for example,
00:40:52 - episodic memory or that kind of property,
00:40:55 - I don't think that they should be left as a last resort, but they should be
00:40:59 - used when they can make new predictions that actually enrich the way you tackle
00:41:04 - experimentally the problem.
00:41:06 - So, for analogical reasoning, in some cases.
00:41:11 - There have been experiments, I wouldn't go now into the details,
00:41:15 - where certain tasks have been solved by animals.
00:41:19 - And the immediate claim has been that this actually is compatible with the animals
00:41:25 - using analogical reasoning here.
00:41:28 - Well, yes, in many cases they are compatible, but they are also compatible with
00:41:32 - other things that we understand better.
00:41:34 - It doesn't mean that they are more likely to be used by the animals,
00:41:38 - but we understand it better. So, I am interested, for example,
00:41:44 - in trying to see whether one could understand the development of these capabilities.
00:41:50 - Because, for instance, when animals resolve novel problems,
00:41:57 - they very clearly resort to some library, some database of experience that they
00:42:02 - have that has been informally acquired through interaction with the different
00:42:06 - affordances of their world.
00:42:08 - And we are testing adult animals.
00:42:11 - We don't know how they acquired that sort of database of relationships.
00:42:15 - And then they come and solve something de novo.
00:42:18 - So how do they do it? Is it because they understand what's going on at the physical
00:42:25 - level or is it simply because they are generalizing from their long-term experience?
00:42:31 - Is but not to so before we now
00:42:34 - solve this this dilemma between let's
00:42:38 - say the decomposition view or the more representation list internal view maybe
00:42:43 - we can look at a bit more the experiments that you have performed and one of
00:42:46 - them which i thought was very interesting was where you compare um the crows
00:42:51 - with with these chaos this is a parrot from from new zealand right,
00:42:57 - where you also actually got some ideas about the enormous individual differences
00:43:02 - and I think that's another aspect we should not lose sight of when we try to
00:43:06 - make some categorical decision on okay they solve it like this or like that because.
00:43:11 - If it's true that there's a large individual variability, it could even happen
00:43:16 - that within the species you cover this whole range and there's not an exclusive
00:43:20 - choice to be made, right? Why not? This is also a possibility.
00:43:24 - So in these experiments, in this comparative experiment in the Krauss and the
00:43:28 - Kess, what did you really observe? What were the key observations there?
00:43:32 - Well, what we did in that particular example that you referred to is offer a
00:43:38 - battery of different tests and trying to see whether the capacity to jump from
00:43:45 - one form of solution to the next and discovering the next one was characteristics of the species.
00:43:51 - And we did find that to some extent.
00:43:54 - But on the other hand, when we examine the reasons for success in certain tasks,
00:44:00 - we found that what you have to attribute to understanding of the physical features
00:44:08 - of the problem could be severely restricted by other differences between the individuals.
00:44:14 - For example, their proclivity to explore the environment in either a tactical
00:44:20 - modality, sorry, a haptic modality, or a visual modality.
00:44:24 - Because certain problems cannot be discovered visually, and others,
00:44:30 - which involve action at a distance, are more difficult to discover by touching alone.
00:44:36 - And differences which one would call non-cognitive get amplified into difference
00:44:43 - in problem solving and tasks which overall are designed for their cognitive interest.
00:44:49 - Because these chaos are more haptic, right, in their interaction with the world.
00:44:52 - Yes. While the crows are a bit more at the distance.
00:44:55 - That's right. At the distance. Yeah. But then, so, in my mind,
00:44:59 - surprisingly, you found these chaos were overall better in this battery of tasks
00:45:03 - than the crows. So what made exactly that distinction?
00:45:08 - Yeah, well, the crows were better at the task that they are well known for,
00:45:16 - which is to use objects to act on target at a distance.
00:45:21 - Basically, to use a tool to retrieve the, although the task was new,
00:45:25 - it had the basic structure of what they are known to be good at.
00:45:28 - But the Chias could do better any task in which direct manipulation of the object
00:45:38 - with their beaks and feet was an advantage.
00:45:44 - And if you were to take naively the results of the experiment in a quantitative
00:45:51 - way, yes, the Chias scored more highly than Euclidean crows.
00:45:56 - But we have left behind, I hope, long ago, this notion of a scalar categorization
00:46:05 - of intelligence between species because they are designed to solve different problems.
00:46:11 - And they basically have different capabilities and motivational differences cause that.
00:46:19 - I should say that you mentioned this is a very important problem when you're
00:46:23 - dealing with intelligent animals.
00:46:25 - I'm using the term here loosely to say one that uses a lot of its own experience to solve problems.
00:46:32 - And almost by very definition from the beginning of the task,
00:46:37 - you are expecting different personal histories,
00:46:40 - different individual histories to accumulate and lead to different adult behavior
00:46:44 - because nobody can guarantee what individuals are going to do.
00:46:50 - So if you look at the diversity of skills in humans and you ask who ends up
00:46:56 - being a professor of Chinese and who ends up being an engineer designing robots,
00:47:02 - for example, or a neuroscientist, it's very difficult to know what was inherently
00:47:06 - different in the alleles of different genes that these individuals had and what
00:47:12 - simply were contingencies in their personal experience that have not been formalized and cannot be.
00:47:18 - The same happens with intelligent animals. They have... Sure.
00:47:21 - But what I found interesting about your interpretation of this difference or
00:47:24 - similarity in the chaos and the cross, where you say, look, they have been specialized
00:47:28 - for different kinds of tasks, so differences should not be over-interpreted.
00:47:32 - But if I look at your result, my interpretation, I could give you an alternative
00:47:37 - interpretation, which actually, they are able to solve the same tasks.
00:47:41 - And in the Simon and Newell view, that might be the same expression of a general
00:47:46 - intelligence, if you want, where anything can become a task.
00:47:49 - However, their performance is sort of modulated by the specifics of their skeletal
00:47:54 - muscle system, their training history, the niche that they sort of have adapted for.
00:48:00 - And for the example you showed where the chaos were pretty good in dealing with
00:48:05 - a little window that was in the index mental box, while the crows could not.
00:48:09 - And the difference being that these chaos, since they're so haptic.
00:48:12 - They chew on everything and they touch everything. So in that way,
00:48:15 - they can discover this window while this crow is just perching somewhere and
00:48:19 - looking down upon the task and never figure it out.
00:48:22 - But that would mean actually that coming from these very different morphologies
00:48:26 - as a bird, being controlled by a fairly similar brain allows them to solve the same task.
00:48:34 - It's just modulated by their physical instantiation, the bodies that they have.
00:48:37 - Is that a reasonable interpretation?
00:48:39 - I think it's perfectly possible. And this is something that we would like to
00:48:44 - tease apart in a whole program of experiments, exactly.
00:48:48 - How much of what we see is just modulation of a very general ability to say,
00:48:55 - for example, represent the problem in some symbolic kind of way that allows
00:49:01 - you to explore virtually different solutions and then implement them.
00:49:07 - And to what extent they don't have this capability. They are very dedicated.
00:49:14 - We know that humans are not completely equally capable of any content-free problems.
00:49:22 - I mean, our ability to solve logical problems. But now there's an interesting
00:49:25 - consequence, right? Because….
00:49:28 - You also used the word intelligence loosely. Yeah. And intelligence is often
00:49:32 - seen as a reasoning capability based on a mysterious g-factor that recently
00:49:40 - Adrian Owen and others have shown might decompose again in other properties
00:49:44 - such as working memory and rule learning and so on.
00:49:47 - But now we have to see that if we want to insist on the notion of intelligence,
00:49:50 - we also have to include morphology.
00:49:53 - So doesn't that actually imply that we might forget about noise of intelligence
00:49:57 - because it starts to become the whole universe well I'm we try sometimes so
00:50:02 - it is very hard to leave it behind for example,
00:50:07 - we did the exercise in our
00:50:10 - laboratory to for different periods to ban certain words and see whether we
00:50:17 - could manage that and one of the words was understanding so could we We go on
00:50:22 - in our lives replacing the word understanding by what we really mean in any way.
00:50:30 - And we fail miserably. We had to bring it back and accept to use it.
00:50:37 - It's the same as doing evolutionary biology without teleology.
00:50:41 - We know that teleology is a shorthand for the action of natural selection and
00:50:46 - the different variations.
00:50:47 - But then if you don't really say what the oak tree tries to do is to make as
00:50:55 - many acorns as it possibly can.
00:50:57 - So you're not making any subjective attribution, but you are using a notion
00:51:01 - of goal, of target, which is for natural selection and not for the individual.
00:51:07 - And the same kind of thing we need to use for animal behavior all the time.
00:51:12 - Yeah, it's also the straitjacket in which behaviorism tried to put itself without much success.
00:51:18 - That's right, yes. It's just, it ends up, it brings you to study only boring
00:51:22 - problems because they are only the problems you can formalize verbally at that time.
00:51:27 - So you have to be a little bit more loose, but at the same time not have the
00:51:32 - illusion that if you say that an animal solves a problem because it understood
00:51:36 - it, you have explained anything.
00:51:37 - Of thing right but now so another animal
00:51:40 - that that you described was was this uh the kakatu this
00:51:43 - indonesian kakatu where actually started to look at
00:51:46 - a very different aspect so first it was more like what kind of problems can
00:51:50 - they solve what kind of tools can they use and actually construct which already
00:51:54 - is amazing but now with this kakatu you went again a step further to say well
00:51:58 - how many steps can they actually chain together to solve a complex problem that they
00:52:06 - actually would never encounter in nature because you built some strange contraption
00:52:09 - that even for humans might be a bit of a challenge initially.
00:52:13 - Where they had to sort of go through different steps to get a food reward.
00:52:18 - So what was the key insight and the key motivation behind that experiment?
00:52:23 - In that particular experiment, the animals had to do a series of actions,
00:52:28 - up to five actions in this case.
00:52:31 - And we'll say actions, they were very different in what actually the motor intervention
00:52:37 - had to be before reaching a target.
00:52:41 - In this case, it was a foot reward. world.
00:52:44 - But the interesting problem was that the sequence could not,
00:52:50 - none of the components of the sequence was reinforced until the whole sequence
00:52:55 - was done and in the right order.
00:52:57 - And that means that the animal could not improve progressively by reinforcement of individual actions.
00:53:06 - But what What it could do is to improve by, in a sense, perceiving the solution
00:53:14 - of anything that shortened the
00:53:15 - chain that was a physical chain of physical devices engaging one another.
00:53:20 - Anything that make it shorter was indeed progress towards achieving the goal.
00:53:26 - And that particular experiment did not require planning as such,
00:53:31 - but it required the capture of experience with a notion of goal-directedness,
00:53:37 - of sort of trying to achieve something.
00:53:40 - But the critical part of that study was that once the animals had learned to
00:53:45 - solve the problem of the sequence, we did controls in which we removed elements internal to the chain.
00:53:53 - And so now the question was, what has the animal learned?
00:53:57 - Is it going to go to the old beginning as required, or is it going to go to
00:54:02 - the first element after the removed one, so that now the problem they face is shorter.
00:54:09 - And what we found is statistical evidence that they can do the latter.
00:54:13 - And that means that somehow, that is, they do the properly functional thing
00:54:18 - of skipping parts of the chain, which are now being rendered irrelevant by the
00:54:24 - transformation of the task. Right.
00:54:26 - And what that means is that the animals are, and now I'm going to use carefully
00:54:32 - my words, the animals are sensitive to the physical interactions between the objects.
00:54:37 - Objects and if i tended to
00:54:40 - say that they understand the physical interaction between the objects but as
00:54:43 - you see i'm avoiding it so they somehow they
00:54:47 - behave as if they could eliminate parts of the task which have become irrelevant
00:54:55 - by a modification that they can see but they haven't experienced before so in
00:55:01 - your mind would be they have some sort of model of
00:55:04 - the overall problem in which they can selectively perform adaptations.
00:55:09 - I think so. But we really, I mean, just being conservative,
00:55:14 - I mean, we don't know how they do it, but we know that we can eliminate some
00:55:18 - classes of explanations which are based on very simple… What is interesting
00:55:23 - here is that from a standard Thorndikean and also reinforcement learning point of view,
00:55:29 - which in general I find a very impoverished way to think about the world and certainly biology.
00:55:34 - You would have to think about a direct reinforcement signal to first identify
00:55:39 - a step in the chain and then to link it together.
00:55:41 - So in your task, that's not possible because I really have to go through the
00:55:46 - whole chain, through all the steps from beginning to the end before I get my
00:55:50 - reward. So then you could argue, well,
00:55:53 - There might be some backward chaining. That means I have a memory representation of this.
00:55:58 - This is one way in which I could talk. I do step A. Step A is committed to a working memory.
00:56:03 - Then I do step B, and step B is inserted in a working memory until I get my
00:56:08 - reward, and then I have all these elements in my memory because they all belong
00:56:12 - together. This is how I can solve the problem.
00:56:14 - Alternatively, you could say, well, maybe Thorndike is still right with this
00:56:17 - law of effect, but the instruction signal is not external. You don't get this
00:56:21 - food reward or the worm or whatever they get, the nut.
00:56:24 - You have like an intrinsic motivational signal. Aha, it's great to solve a problem.
00:56:30 - And that's your rewarding signal that helps you to glue the steps of the chain
00:56:33 - together. So which of these two interpretations have your difference?
00:56:38 - I'm happier with the second. I would imagine that I don't think we could make
00:56:41 - the first one work in this particular task.
00:56:45 - Because the animals cannot solve the problem backwards and they have no experience.
00:56:50 - Experience the final elements of the task at all at the time that they start
00:56:55 - working on the other side of the chain.
00:56:56 - And so a backwards reinforcement procedure would never take them to the end.
00:57:02 - So, but somehow you have to give the system a method to progress,
00:57:09 - even if it's modeling it mentally, to know what is an improvement and then build on that.
00:57:16 - You may call that reinforcement or virtual reinforcement, if you want.
00:57:20 - It doesn't involve the physical.
00:57:22 - But that's interesting, right? Because hidden in that, you have a definition
00:57:26 - or a notion of, let's say, subtask completion, right?
00:57:32 - There must be something very identifiable about the event so that they can drive this intrinsic signal.
00:57:39 - So what could that be? Well, in the case of this particular task,
00:57:45 - where you have different physical devices engaging one another,
00:57:49 - if you imagine an image representation of that in your mind,
00:57:55 - you can see that you could visualize the downstream device as being movable when something,
00:58:05 - you know, a poke in the wheel has been removed.
00:58:08 - And so if that is blocking it, if I remove it, then the wheel can turn,
00:58:12 - something like that. So you need to have a notion of that sort of physical interaction.
00:58:21 - So you're saying, this is actually quite a strong assumption,
00:58:24 - which is very interesting, because you're saying, well, actually the animal
00:58:28 - has to have some sort of understanding of the whole task,
00:58:34 - of this whole machine that it has to deal with.
00:58:37 - And then within that model, it says, aha, I did step A now.
00:58:41 - Well, it's important that in this particular task, I don't want to over-interpret
00:58:45 - the data or even give them that impression at all.
00:58:49 - In this particular test, they could progress by rattling at random everything.
00:58:57 - But whenever they achieve a movement, they have to remember it perfectly the next time to it.
00:59:07 - And so, and that way, like a ratchet, they get closer and closer so that one
00:59:12 - day they come in and do one A, B, C, D, E, and then they get the fourth.
00:59:16 - So it doesn't require that they
00:59:19 - understand the physics of it to reach the solution for the first time.
00:59:24 - However, when you modify the task, you transform it by altering the order of
00:59:31 - things or removing one thing.
00:59:33 - If that was the way they learned it,
00:59:36 - and that was all that is stored in the mind of the animal, then it would not
00:59:42 - go zoom in what is the right movement
00:59:46 - now after the transformation of the task. And that's what they do.
00:59:51 - And the implication of this is that they may achieve it by something which is
00:59:56 - not very systematic and based on understanding and the logic of physical interactions.
01:00:01 - Actions but once they have done it they are sensitive to the actual physical
01:00:07 - needs of the task to create a novel solution the next time over but imagine
01:00:13 - i build a machine where i have the same.
01:00:16 - Elements i have this this this bolt i have to turn and so on but the linking
01:00:22 - to the next step which doesn't say something they have to unplug or something
01:00:27 - has to pull is not mechanically identifiable.
01:00:30 - I do that in the background through a computer, let's say.
01:00:33 - But the order stays the same.
01:00:36 - Would you believe the cockatoo would be able to solve that problem equally well
01:00:41 - if it cannot recognize the physical linking, the mechanical linking of the steps?
01:00:48 - Um...
01:00:49 - The true answer is I don't know, but I don't think is,
01:00:54 - I think it connects with some experiments that have been done in chimps, in children,
01:01:01 - and to some extent also in birds, in which you create, you rig up devices so
01:01:09 - that the movements that you cause are physically intuitive and logical or not.
01:01:14 - And you see whether the animal has greater difficulties when
01:01:18 - the action is not what you
01:01:21 - might expect from normal physical laws so we
01:01:24 - can do some of that and we indeed we we
01:01:27 - try but i can't at the moment answer your thing maybe
01:01:30 - you should build a magical machine for the bird to take
01:01:33 - apart that's right yeah to some
01:01:36 - extent your task is complicated because there
01:01:40 - are physical manipulation aspects to it
01:01:43 - which are quite challenging for for birds um but
01:01:47 - listening to you describe it it reminded me of some you know cognitive tasks
01:01:51 - that for instance development psychologists like piaget developed his serial
01:01:55 - order task which was to take a pile of sticks of different length and and lay
01:02:00 - them out in order of ascending size and then many people have looked to that in child development,
01:02:06 - or transitive inference tasks that have been investigated in children, also in chimps.
01:02:14 - Is it, I mean, have these been investigated with birds as well,
01:02:18 - these kinds of more cognitive tasks that don't give them that physical challenge so much?
01:02:22 - Yes, yes, they have. For example, the work by Alan Camille and his colleagues
01:02:30 - on what you mentioned of transitive inference is very interesting because what they show is that,
01:02:41 - some kind of physical transitive inferences required to solve some tasks.
01:02:48 - For example, if you learn that A, I'm going to try to reconstruct now,
01:02:54 - if you have that A is bigger than, I don't want to do a misrepresentation of the task.
01:03:03 - The point I'm raising is when you compare species which have a complex hierarchical society.
01:03:09 - And some which have a more egalitarian one, what you find is that they can transfer
01:03:18 - that skill to tasks which have the same logical structure, but different content.
01:03:23 - Yeah, yeah. So rather than having a modular device just capable of solving the
01:03:30 - social problem they have, they have the general, they have developed that.
01:03:34 - Now, have they developed this because they've experienced it repeatedly and
01:03:40 - they just abstract through their history this problem that the others have not?
01:03:45 - Or do they have a pre-programmed inherited logical module that allows them to
01:03:51 - do that kind of inference?
01:03:55 - That we need to do developmental studies to test.
01:03:58 - You have to see what happens if you raise animals with different level of complexity
01:04:02 - and see what kind of logical tasks they can solve later.
01:04:06 - So now you did say that these crows use tools and cockatoos do not in the wild.
01:04:17 - Can you really be sure about that?
01:04:19 - No. Okay. I said that crows are very well known for their extensive use of tools
01:04:26 - in the wild and cockatoos are not known to use tools.
01:04:29 - Now, if they were as intense tool users as the crows are, we would know.
01:04:36 - But they have not been studied in the wild, actually, in greater detail.
01:04:41 - So there is, I would never claim lack of an ability or someone is going to find it.
01:04:48 - If they have the skill, why wouldn't they do it?
01:04:51 - In the case of capuchin monkeys, they were shown for many years to have great
01:04:56 - capability for the use of tools in the laboratory,
01:05:00 - and people kept repeating that they don't do it in the wild until some groups
01:05:05 - of researchers found them doing it in the wild and doing it a very sophisticated
01:05:09 - kind of tool, which is now a classic study.
01:05:12 - So in your torture conclusions at some
01:05:15 - point you made the point we play chess because we
01:05:18 - are bad at it so what are you trying to this is an intriguing statement the
01:05:24 - point I was trying to make is not that we play chess I was saying we are impressed
01:05:29 - by performance in chess and people use chess as I think this idea is borrowed from Chomsky,
01:05:37 - it's not my own but I just.
01:05:40 - I can't locate at the moment where I read it, but I remember it labeled clearly
01:05:44 - as coming from reading that.
01:05:46 - Basically, if you compare something like the linguistic skills of humans and
01:05:52 - the speed with which we acquired it at an early age with extreme poverty of data,
01:05:58 - although not absence of data, as it has sometimes been claimed, but poverty of data.
01:06:02 - We very quickly learn extraordinary complex tasks, while chess has a very limited set of rules.
01:06:11 - And we nevertheless don't become half as good at doing it.
01:06:16 - And there are great individual differences in chess playing ability while there
01:06:21 - are perhaps less differences or we don't use them as a measure of capability
01:06:28 - for producing normal human speech.
01:06:32 - So, all I was saying is that we have to be careful that we may end up using
01:06:39 - things which are particularly outside the range of competencies of a species as interesting,
01:06:49 - particularly because the species is bad at it.
01:06:52 - Right. Although not everyone is a poet.
01:06:55 - Not everyone is a poet, indeed. So then you made another important point in
01:07:02 - your conclusions, which had a lot to do with sort of the comparative aspect
01:07:05 - of this, like what the things we learn about birds,
01:07:08 - how would it generalize to other species?
01:07:11 - And you made this point about the relationship between body size and brain weight
01:07:15 - that you found sort of speaking to this point.
01:07:18 - So what's the message there? Yes, the point I was trying to make is we know
01:07:23 - that absolute brain size is not a very rich indicator of capability.
01:07:28 - But on the other hand, it's hard to believe that the absolute size is entirely irrelevant of a brain.
01:07:36 - But basically, by people playing around with different ways of plotting brains,
01:07:45 - different sizes of brains, scale them in different forms, they found a way in
01:07:51 - which humans do particularly well.
01:07:53 - And that has stuck. And this is actually to look at the relation between human
01:07:59 - brain size relative to what you would expect from a mammal of that size.
01:08:04 - And you find that the residual on that overall correlation favors humans very dramatically.
01:08:09 - We don't have the biggest brains in nature, but we have the biggest residual
01:08:14 - with respect to animals of our size.
01:08:16 - All I was pointing out is that if you look at the whole of birds,
01:08:20 - and particularly if you looked at passerines, The,
01:08:24 - area in this plot of brain size versus body mass that you have is very similar.
01:08:31 - So birds tend to be smaller than at least the larger mammals,
01:08:35 - but for a given size, they have similarly sized brains.
01:08:39 - But within that, what you find is that the parrots and the crows,
01:08:44 - which happen to be the groups of animals for which we have greater evidence
01:08:48 - of these cognitively difficult tasks are also the ones which have greatest residual
01:08:56 - respect to the overall regression in the birds.
01:08:58 - So all I was saying is that, yes, there are differences between species.
01:09:04 - Not everybody is equally smart.
01:09:06 - Maybe parrots and crows are really capable of greater general logical processing
01:09:15 - of the the world and they have something more of a G-factor than,
01:09:19 - others, but as a biologist, I wouldn't jump to that until I exclude the direct
01:09:25 - link with their ecology.
01:09:27 - Right. But now, with birds, also crows, they're one of the few animal species
01:09:31 - where people have shown that they do have something like self-consciousness, right?
01:09:36 - So you can stick, it's a typical task, right?
01:09:39 - You stick something to the body, you put the animal in front of a mirror,
01:09:42 - and apparently they will take it off. So.
01:09:46 - Do you buy these kind of experiments?
01:09:49 - Do you think that something like self-consciousness is important in the kind
01:09:53 - of mental operations we're talking about for these tasks, or you see this as
01:09:57 - completely irrelevant? relevant?
01:09:58 - I buy the experiments in terms of empirical results.
01:10:03 - What I don't think is fair is to jump from the ability to address behavior to
01:10:09 - your own body on the basis of external stimuli to some philosophical notion
01:10:13 - of selfness and identity.
01:10:16 - You could be trained by watching mirrors to identify self by the contingency
01:10:23 - between your movements and the stimulus that you're seeing outside.
01:10:26 - That's a little bit more difficult than looking at your own hand.
01:10:30 - You are looking at your hand in the mirror, but still you are learning from your experience.
01:10:37 - But that seems an interesting transition because with the problem-solving behavior,
01:10:41 - you were very much at the end of, okay, there's a complex internal model in
01:10:45 - which you perform operations.
01:10:47 - Well, if we now talk about self, you seem to try to reduce it away to simple
01:10:52 - externalized stereotype behaviors?
01:10:54 - Why is there not a model of the self then as well, as much as there is one of a complex task?
01:10:58 - Well, Paul, if you are insinuating that I'm contradictory, I'm quite willing to accept it.
01:11:07 - I can't help but reflecting the different realities that we observe,
01:11:16 - having a different degree of confidence in what we can infer.
01:11:21 - I don't think I think it's fair to describe my previous statements and saying that I was...
01:11:28 - Kind of defending a highfalutin interpretations whenever a simple one would do or the opposite.
01:11:34 - No, no, I wasn't trying to say that. Closer to the culture one,
01:11:38 - but often we fail and then we have to elaborate on that.
01:11:41 - My key point was in the problem solving, it's clear that you do assume that
01:11:46 - there's an internal model, which is completely reasonable.
01:11:48 - It's an internal model of the task. While in the self-oriented behaviors,
01:11:53 - you might equally say there is an internal model of self.
01:11:56 - Why not? But you seem to sort of not want to go that way.
01:12:00 - I'm hesitant because I don't find the evidence of the mirror self-recognition
01:12:08 - task sufficiently compelling to tell me that the animal has to have what we
01:12:14 - normally understand by the notion of the self, the notion of agency.
01:12:17 - I believe that it's natural that the club of those that can do it can only grow. It can't get smaller.
01:12:28 - So people start by showing that it only happens in humans, and they say that this defines humans.
01:12:34 - Then they find that this also happens in chimps, and then next they find that
01:12:38 - it also happens in some corvids, in magpies.
01:12:42 - And then they keep up, and then eventually they show that it works in fish.
01:12:47 - And now they have a problem.
01:12:49 - Do fish also have this virtue that we call identity, self-conscious, self-recognition?
01:12:59 - Or the experiment didn't require that in the first place when we saw it in humans.
01:13:05 - This is a very common process. Sure. So, Alex, to finish up, two questions.
01:13:12 - It's clear that you have a broad experience in this domain, studying animal
01:13:17 - behavior over many years, having deep insights in the capabilities of many of these animal species.
01:13:24 - So, what would be Alex's law that we should adhere to in the study of animal cognition?
01:13:33 - Hmm.
01:13:35 - I would say that to, I'm very Tim Bergen in this,
01:13:41 - and, you know, I think that the advice of, or the way that Nico Timbergen in the 60s, in 63,
01:13:51 - structured the program of ethology is still a very good program, you know.
01:13:55 - He said, when you look at any behavior, but you could apply to any biological
01:14:01 - trait, don't think that a single level of explanation is sufficient.
01:14:06 - So don't use exclusively mechanistic interpretations.
01:14:09 - This is the way they do it and satisfy with that, nor purely normative ones.
01:14:15 - This is why they do it this way and satisfy with that. I think an interaction
01:14:20 - between normative and mechanistic approaches is absolutely essential.
01:14:27 - But another thing that is essential is to use the right level of reductionism.
01:14:32 - If we were to jump now immediately with our current level of knowledge to some
01:14:39 - of the very basic properties of birds'
01:14:43 - nervous systems, I don't think we would make much progress in the kind of problems
01:14:48 - we are facing at the moment.
01:14:49 - We can still do a lot by working at the behavior of the animals and making inferences
01:14:56 - from that, taking the ecology into account, taking evolution into account.
01:14:59 - But of course, in the end, you do want to reduce one step at a time and go as
01:15:05 - far as possible to the basic machinery of how the animals achieve it.
01:15:09 - So we should follow some plurality, though, in our view on these phenomena.
01:15:13 - Absolutely. And there's one thing I said right at the end in the conversation
01:15:17 - in the question time was that I believe much more in multidisciplinarity than
01:15:21 - I believe in interdisciplinarity.
01:15:23 - What I mean by this is we can't be equally good at everything we do,
01:15:28 - but we can meet with colleagues and form teams where everybody is very good at everything.
01:15:35 - A particular way of looking at nature. And in that, you also see a clear role
01:15:39 - for computational and robotics-oriented approaches.
01:15:43 - Very much so. I really think that we are,
01:15:48 - well, not fully because here we are doing it,
01:15:52 - but we would be losing and missing opportunities if we didn't use the wisdom
01:15:58 - acquired by people in AI and robotics on what you need to have in a machine
01:16:04 - to be able to take autonomous decisions and to construct novelty in behavior.
01:16:09 - And we will not use that wisdom to interpret the behavior of animals that we
01:16:13 - see doing that, but we have no idea of what process is underlying it.
01:16:18 - So I think working together would tell robots, sorry, roboticists,
01:16:24 - what kind of problems they may aspire at solving by looking at what animals can do.
01:16:30 - And it would tell us how some of these problems can actually be tackled because
01:16:35 - they have been tackled in machines being built.
01:16:39 - So now five years from now, Tony and I will come visit you in Oxford and we're
01:16:43 - going to confront you with the prediction you're going to make today.
01:16:46 - So I'm going to ask you whether it was validated or invalidated in this intervening period.
01:16:51 - So what's the one prediction you really would like to make today and you really
01:16:56 - would like to stick to and rigorously investigated over the coming five years.
01:17:01 - Gosh, Paul, you didn't warn me of this question.
01:17:05 - Surprise, surprise. Yeah, yeah, yeah. Give me a few seconds to say something
01:17:12 - non-completely trivial. Um...
01:17:18 - I would say that certainly in my own field,
01:17:23 - if I can be a little parochial, the behavioral ecology field that evolved out
01:17:28 - of ethology by moving the swing of interpretation towards purely functional
01:17:34 - analysis would have moved considerably back in the direction of paying attention
01:17:39 - to how animals do things.
01:17:42 - But this is a prediction about science, not about animals, so I may be cheating a little bit.
01:17:49 - But I'm saying that what is going to happen is that people are going to realize,
01:17:54 - that making purely normative functional interpretations without looking at how
01:18:00 - animals actually achieve it becomes sterile in the end. So we need that.
01:18:06 - Wonderful. Alex Czelnik, thank you very much for this conversation.
01:18:09 - Thank you. It was a pleasure.
01:18:11 - Music.
01:18:39 - Go to csnnetwork.com.
01:18:43 - Music.

CSN Podcasts

Alex Kacelnik on new caledonian crow and tool use

Season 2013

Description

CSN Podcasts

Episodes

Etienne Koechlin on prefrontal cortex and cognitive control

Eberhard Fetz on brain-computer interface and neurochip

Tim Pierce on olfaction and chemical sensing

Marc Toussaint on planning as inference and graphical models

Cyriel Pennartz on hippocampus and ventral striatum

Timestamp

Be the first to leave a comment

Leave a commentCancel

Legal

Follow us on

Alex Kacelnik on new caledonian crow and tool use

Description

Episodes

Timestamp

Be the first to leave a comment

Leave a commentCancel

Tags

Legal

Follow us on

Login to enjoy full advantages

Please login or subscribe to continue.

Go Premium!

Enjoy the full advantage of the premium access.

Stop following

Cancel subscription