The concept of instinct lays exactly in the center of my scientific interests.
It lays here largely because I was naively assuming that is dead for quite a long time – and is there a bigger pleasure than a dissection of dead ideas? Instinct was declared dead many times and for many reasons – that it lacks a clear definition or doesn’t explain anything. But it always reappears, this way or another, just like a monster from a horror movie. And it happened again, this time in debates on AI.
As you may know, AI is achieving amazing things nowadays – it composes some mediocre pseudo-Baroque music, probably outperforms pigeons in detecting breast cancer and is paving a way towards antyutopian, totalitarian hell where everybody is tracked and controlled, just like in some Orwell’s books that I loved when I was a teenager, before I have fortunately grown up*.
AI achieves all these wonders by learning. You show it thousands of pics, train for a few days, emitting more CO2 than Baltic republics in a year, and it will learn how to recognize a potato. The efficiency of those algorithms has led some people to believe that is just enough to let algorithms learn to achieve the Artificial Intelligence at some point. This hope – or threat – is being criticised with the use of none other, but the instinct, the the zombie of concepts, the undead nightmare of behavioral sciences.
The charge is simple – as Antonhy Zador Anthony Zador wrote recently, there are many things that animals do not need to learn. Examples follow:
A squirrel can jump from tree to tree within months of birth, a colt can walk within hours, and spiders are born ready to hunt. Examples like these suggest that the challenge may exceed the capacities of even the cleverest unsupervised algorithms.
His point is clear – we probably should add some innate constraints to our algorithms to achieve the level of intelligence of a mouse, for example, or just to make them more efficient. Animals are not blank slates, maybe neither should our artificial intelligencies. If it is right or wrong I cannot tell, for the last time I used the neural network the result was the rear part of the mouse got recognized as its head.
But what is interesting for me, a simple-minded behavioral scientist, are terms used in the paper. To be fair, Zador does not use the term instinct. What he does is to mention behaviors that are innate. And as innate he will assume behaviors that are created by innate mechanisms that are encoded in te genome; behaviors that are not learned (colt’s walking); behaviors that are present from birth (spider’s hunting), behaviors that develop without tutoring and that are species-specific (making different burrows different Peromyscus species). All those behaviors are described in ethological literature as instinctive – and, as you will see, this richness of meanings will turn out to be problematic.
A Twitter debate that happened a propos a similar discussion on AI by Gary Marcus and Yoshua Bengio was actually also a debate about the instincts and the role of genes, experiences and learning in the forming of animal behavior. Yes, nature – nurture debate is back, my dears. The existence of instincts is now an argument for developing our algorithms in a specific way.
We understand the concept of instinct almost instinctively, and we usually just assume that it is clear and well defined. But what does it really mean that the behavior is instinctive? Does it really have a clear meaning? And should we use it in our scientific debates, either on animal behavior or AI?
The concept of instinct – in this or another form – is very ancient. I am not a historian of ideas, so I will not naively try to trace its beginnings – but let me just say that the modern concept of instinct was formulated in the XX century by Konrad Lorenz. In his 1932 paper, Methods of Identification of Species-Specific Drive Activities in Birds, he says that the behavior is instinctive if it satisfies one or more of five criteria: 1) if it appears in an animal reared in isolation, without any tutoring, 2) if it is performed in a stereotyped way by all individuals of a specie, 3) if there is a striking mismatch between the typical intellectual abilities of an animal and the abilities that it wouldhave to posses to solve a given problem by insight (e.g. relatively stupid bees are able to show their companions where flowers are by dancing), 4) if the behavior can be elicited in an innapropriate context, suggesting that is not performed consciously, 5) if the behavior is performed in a stereotypical way even in context that is dfferent from the one in which it originally evolved. Niko Tinbergen, who together with Lorenz whole field of ethology, described instinct also as highly stereotyped, coordinated movements, the neuromotor apparatus of which belongs […] to the hereditary constitution of the animal.
Yes, we have seen those different meanings of instinct in Zador’s paper and in the Twitter discussion. And Lorenz’ and Tinbergen’s definitions are not the only ones. Patrick Bateson in his critique of Steven Pinker’s book on innateness lists additional meanings:
“Apart from its colloquial uses, the term instinct has at least nine scientific meanings: present at birth (or at a particular stage of development), not learned, developed before it can be used, unchanged once developed, shared by all members of the species (or at least of the same sex and age), organized into a distinct behavioral system (such as foraging), served by a distinct neural module, adapted during evolution, and differences among individuals that are due to their possession of different gene”
I think we can add a few others. People often say that instinctive behaviors are genetically preprogrammed; that they are somehow hardwired in the brain. Sometimes scientists also automatically assume that behaviors crucial to survival – foraging, mating and fear – are instinctive by definition. Instinctive behaviors are sometimes told to be biological, as contrasted with psychological; some people would also probably say that ancient behaviors, ones present in so called lower animals, are also instinctive.
At this point, you might say: well, and so what? Behaviors that are genetically determined do not require learning and are hardwired; they are also species specific and fulfill other criteria as well. So it is absolutely normal that we have different meanings: instinctive behaviors just have many characteristics! And indeed, people very often use different meanings of instinct almost interchangeably; Steven Pinker, for example, does not define it at any point in his whole book on innateness – he just assumes that we know what he means.
But is it true? Do so called instinctive behaviors posess all those characeristics? Are different definitions compatible with each other? Let’s have a look at a few examples.
Let’s start with an easy example and take two meanings of instinct: present from birth and not learned. The two features seem to fit each other perfectly – but is it so? Can animals learn anything before they are born?
The answer is yes; there are plenty of behaviors that animals learn in utero. Superb fairy wrens learn the specific call from their mother while still in an egg, which probably constitutes an adaptation against brood parasites. When ringed salamander eggs are exposed to chemical cues from predator mixed with a neutral chemical, they develop alarm response to that chemical after hatching; in humans, prenatal experience with speech heard in utero influences infants’ perception of speech and maybe even shape their cry patterns after birth.
All those behaviors are present at birth – and yet all of them are learned! It seems that not all meanings of instinct always peacefully coexist. But let’s take another example – drinking water when thirsty. It is behavior that is crucial to survival; it is obviously present in all individuals of a given specie and is clearly adaptive. But well, it is apparently learned, at least in rats. Rats must learn the association between dehydration and the relief from dehydration achieved thanks to drinking water – without this experience, they will not seek water when thirsty. If you feed them with a liquid food that does not allow them to develop dehydration and then – at some point – dehydrate them artificially by the injection of salt, they will not increase their consumption of water.
As Bateson mentioned, one of the meanings assumes that in case of instincts differences among individuals are due to their possession of different genes. What is interesting, there are behaviors that differ between animals that are otherwise clones, having an identical set of genes. Among identical genetically pea aphids, you can observe differences in they startle behavior – when presented with a loming stimulus simulating a predator, some of them will jump out of the leaf, while others remain feeding. Similarly, individual fruit flies may differ in their thermal preferences. Those differences are stable within individuals. It is true also for bacteria – they exhibit surprisng phenotypic variability even without variability in genotypes.
What is the source of those differences if they cannot be explained by genes? It can be caused by different experiences, but It seems that the development of an organism is stochastic – it is not an execution of a program and even with a smilar starting condition can give variable results; levels of gene expressions in cells may differ due to purely random processes. Positive feedback – e.g. when a gene’s product might enhance gene’s expression – can amplify those random fluctuations, leading to different outcomes. What do we have here are behaviors that are present from birth, not learned, unchanged once developed – but differences between individuals cannot be explained by the difference in genes, though they are for sure biological in origin.
Aforementioned variability brings us to another set of definitions that may not be always compatible – hardwired and highly stereotyped.
I am never entirely sure what does it mean when it comes to neural circuits, but the hardwired circuit is probably circuit that is not variable, that is set very precisely during development by some genetic instructions and that controls behavior. The problem is that neural circuits that produce invariant, highly stereotyped outcome are very often themselves variable. The trength of connections between neurons that control heartbeat in leech vary between individuals, even though the outcome is identical. As authors write in the introduction, each animal arrives at a unique solution for how the network produces functional output.
The same can be said about famous crustacean stomatogastric ganglion – those neurons can produce the same pattern of activity even though their connectivity may look differently in different individuals. Similarly, in Drosophila olfactory system interneurons show interindindividual variability in physiological properties, neurotransmitter profile and connectivity**.
What shouldn’t be surprising. Individuals are different: they differ in size, strength; they may develop in different environments. The nervous system cannot have one pre-specified, rigid solutions to all problems; it must find a solution that will work in specific circumstances: it must be robust. It reacts to what happens during the development: if you artificially enhance spike production in a neuron in a developing embryo, the cell will respond by decreasing the expression of excitatory and increasing the expression of inhibitory transmitters to keep the proper level of excitation. The resulting variability in connections or ion channel expression is probably a way to achieve consistency in behavior. You can read more about the topic in a paper that Robin Hiesinger and Bassem Hassam wrote on variability and robustness.
Another example will show us that behaviors that are fundamental to survival, ancient and controlled by highly conserved brain region may show unexpected plasticity at the neural level.
In a 2017 article by Ryan Remedios and Ann Kennedy authors imaged neurons in the ventrolateral subdivision of the hypothalamus (VMHvl), a structure that is related to mating and fighting, behaviors expected to be instinctive, at least by the authors of the paper. But they show an interesting thing: when an experienced male mouse interacts with an intruder that is either male or female, separate neuronal populations are activated depending on the sex of the intruder. But in case of an inexperienced male, those populations are overlapping and separate only gradually with sexual and social experience. As the claim at the end of their summary:
More generally, [these observations] reveal plasticity and dynamic coding in an evolutionarily ancient deep subcortical structure that is traditionally viewed as a “hard-wired” system.
Finally, let’s look at behaviors that are present in all individuals – so called universals. The universality of a given behavior is often taken as an indication of its innateness; they are viewed as genetically determined and not learned – how can you expect all animals to develop a behavior if its development would depend on experiences that may differ?
Well, there are experiences that are universal, things that happen to all individual of specie which can be reliably used as a source of information that will guide the development of a behavior that is also universal.
For example, all chickens will look at their feet. And it may help them to learn to properly rrecognize worms – yes, small chicken learn to recognize worms by looking on their own feet. If you make them unable to see them by making them wear sockets, they will be much eager to pick a worm when they will spot it for the first time in their lives.
Small mallard ducks follow the calls of their mothers just after hatching – they seem to have an innate recognition of a species – typical call, as they prefer it over calls of other species. But it turns out thet it is learned. Before hatching, when a duck breaks into an air bubble within an egg, it starts to produce its own vocalisations. Gilbet Gottlieb has shown that they learn to recognize calls typical for their species’ mothers by listening to their own vocalisations that share some similarities. If you devocalize ducklings hile they are still in an egg, they will be unable to recognize it.
Listening to one’s own voice is also an experience, experience that is universal and will reliably occur in an every generation – just like a possession of a certain gene.
As you may see at this point, different meanings of instinct may not really compatible with each other; closer inspection shows that it is an incoherent whole. And this incoherence may lead us astray, it may make us surprised by discoveries that are not at all surprising.
This is exactly what happened to the authors of the aforementioned hypothalamus paper. They start with a sentence: all animals possess a repertoire of innate (or instinctive) behaviors, which can be performed without training. And a moment later they write: here we report that hypothalamic neural ensemble representations underlying innate social behaviors are shaped by social experience, just to finish their summary with already quoted fragment, that their results reveal plasticity and dynamic coding in an evolutionarily ancient deep subcortical structure that is traditionally viewed as a “hard-wired” system.
Authors seem to be puzzled by those inconsistencies – we have an instinctive behavior that is not learned by definition, and yet an ancient structure that controls this behavior is shaped by experience! They openly ask: how can these findings be reconciled with the “innate” nature of mating and aggression?
They try to give some answers – maybe here are downstream areas that are truly hardwired! We just need to find them!
But maybe we don’t really need to? Maybe this whole result is shocking just because we never thought about our hidden assumptions? That an innate behavior must be unlearned and hardwired? Well, in light of what I wanted to show you it absolutely doesn’t need to. We are again misled by our concepts, by labels that have huge historical luggage of meaning and that are not really useful now.
But what should we do? Is really the concept of instinct not clear enough to be useful? Should we drop it? I think yes – it was around for too long, acquiring too many meanings on the way through centuries. Even if we would cleary define it whenever we use it, other meanings may appear anyway in the minds of our readers. When we say unlearned, they will read inborn or universal, what may not be a case.
Should we then coin a new term? New terms can be fun, but they are rarely adopted by people. Biology and other science are full of terms that nobody uses, or uses them in a way completely diferent from intended, just like Dawkins’ memes.
Fortunately, don’t need to do it. We have a term that already exists and is often used to describe many of behaviors we mentioned here. What is more important, this term is free of all assumptions that we associate with instinct – this term is robust behavior. People use it when speaking of behaviors that are stable, occur reliably in a laboratory environment, that may not require too much learning to develop.
If I would like to specify the meaning a little bit more, I would say that robust behaviors are the ones that you can find in most of the individuals of a given specie. They develop in spite of developmental perturbancies; they are usually highly stereotyped. But they can be shaped by experience; they are controlled by specific neural circuits, that can anyway vary between individuals (and sometimes, who knows, might be even controlled by different circuits in different individuals!). Robust behavior does not need to fulfill all those definitions that we attach to the concept of instinct. This concept is just a nice, open and tolerant guy who does not exclude anyone just because they do not obey some artbitrary rules. It gives us freedom, and frees us from artificial surprises that instinct generates, whenever we find unlearned beavior that is not hardwired or vice versa.
This way of thinking is absolutely not new – the whole developmental sytems theory goes more or less on similar lines, although for me it is still slightly too eclectic. There is a wonderful book on the critique of instinct by Mark Blumberg. But as I said at the beginning – instinct is very hard to kill. It will be still hanging around, whatever we do, just because it’s ancient, and this post will of course not change that.
But don’t tell me I didn’t try!
* in fact it was only my wife who convinced me that they are so bad, thanks!
** Wait, wait – you may say. But the world is full of examples of neural circuits that are very invariable! And that’s true – fly eye is an example of a very precise and repeatable wiring. The same for Caenorhabitis elegans , when development seem just like a very precise execution of a program, with all hermaphroditic worms having exactly 302 neurons. But even in those cases invariable outcome is a result of a stochastic process – in both cases during the development cells are going trough a selection process; some will become neurons, some will die – but you cannot predict which ones will survive (look here and here).