Do Animals Have Explanatory Knowledge?

In three previous posts (here, here, and here) we talked about animal learning, particularly classical conditioning. However, some animals display complex abilities to learn that seem to exceed the sort of simple association learning that classical conditioning is capable of. I gave one example of this with the existence of animal teaching and memes.

One question that sometimes comes up is if animals are capable of learning anything like unto explanations?

It’s a given that animals don’t have a universal ability to create explanations and thus are not universal explainers. But animals do sometimes display behavior that leaves us with the impression that they understand things in the form of simple causal explanations. What other alternative explanations exist that might also explain this behavior?

One possible way to explain it is something similar to Computer Reinforcement Learning. This is a mathematical theory that also seems like it’s learning simple causal explanations but really isn’t. However, the existing theory of Reinforcement Learning can’t fully explain animal learning yet because it requires thousands or millions of examples to work mathematically. But the existence of the theory does at least suggest that some things that seem like causal explanations aren’t.

Furthermore, David Deutsch raised this very issue in Beginning of Infinity. He gave the example of an ape watching another ape cracking a nut open to eat the seed inside and then doing it themselves. This seems like it requires the ape watching the behavior to understand the causal connection between cracking the nut and getting to eat before it learns to do it itself. Further, an ape cracking a nut is clearly not simply repeating preprogrammed behavior. So this complex memetic learning behavior of apes initially forced us to conjecture that animals had some explanatory knowledge.

However, Deutsch points to the work of Richard Byrne’s on Behavioral Parsing. This theory suggests that imitative animal behavior that seems like it requires explanations may be simple parsing that requires no explanatory knowledge. (BoI, p. 407) [1]

However, it’s unclear if Behavioral Parsing can actually explain all animal behavior that seems to require explanatory knowledge. This post will document a few such cases.

Doggy Rides: A Common Example

My sister’s dog – typical of dogs – gets very anxious and excited if she sees my sister and mother putting on their coats and heading out to the garage where the car is. She’ll whine and try to get in your face to demand she is brought along for the ride. If you tell the dog something like “If you stop whining, we’ll take you along for the ride” the dog will suddenly be seen to putting great effort into not whining out loud anymore while still visibly being anxious and excited. But the dog will refuse to stop whining if not promised to come along.

Behavior like this is complicated enough that it can’t be explained easily via simple mechanical associations. Current Reinforcement Learning theory could not explain it either because it would mathematically require millions of examples. Nor is there any sort of imitative behavior going on that behavioral parsing can be used to explain. So it at least seems like the dog has conceptual understandings in her mind that involve causal understandings such as:

  • If you put on a coat and head to the garage, that implies the owner will be riding in the car — possibly without the dog.
  • But if the owner puts on a coat and heads to the front door, that’s just a walk which doesn’t last as long and isn’t as bad.
  • But if the dog begs to come by whining, she might be brought along for the ride and not left alone in the house.
  • If the dog is told that if she stops whining she will be allowed to come along, she should make an effort to control her audible sounds, even though it’s hard to exhibit such self control, so that she will be allowed to come along for the ride.

These simple ‘causal understandings’ that the dog exhibits function equivalently to a set of simple ‘explanations’ that involve the dog being able to understand how events will unfold over time and how her actions might intervene with the outcome of those events.  

As with all theories and conjectures, the question is really “Do you find that you can’t explain the animal’s behavior without referencing something like a causal explanation?”

Squirrel Intelligence?

Roger Penrose argued that even lower animals show adaptive learning that suggests a form of intelligence. Citing a television program:

I was particularly struck by a sequence in which a squirrel realized that by biting through the wire along which it was crawling it could release [a] container of nuts suspended at some horizontal distance away. It is hard to see how this insight could have been instinctive or any part of the squirrel’s previous experience. To appreciate this positive consequence of its actions, the squirrel must have had some rudimentary understanding of the topology that is involved. It seems to me that this was an act of genuine imagination on the part of the squirrel—which surely requires consciousness!

(Shadows of the Mind, p. 408. Emphasis mine.)

I’m not advocating for Penrose’s choice of the word “consciousness” which I have argued is a poor choice of word due to its vagueness. But this example does have the characteristics we’re looking for, which I emphasized in italics. It’s hard to see how squirrels could have instincts for chewing through wires to get food. And for this particular situation, it’s even hard to see how it could have previously, by chance, learned that chewing through wires resulted in food. And there is no imitative behavior here. So Penrose finds himself conjecturing that the squirrel intelligently conceptualized what would happen causally should it chew through the wire using something like a causal explanation.

Now, of course, this isn’t a well-controlled test case. It’s at least a possibility that this squirrel had somehow learned this previously, just by chance, and was applying its previously learned behavior.  

So how might we go about setting up a test case that would eliminate that possibility?

Animal Causal Conceptualization

Nicholas Christakis documents a number of experiments meant to try to work out something similar to this question:

One experiment placed pairs of capuchin monkeys in adjacent sections of a single test chamber, separated only by a mesh partition. In front of each of the monkeys was a cup holding food and a bar that, when pulled, would move the two cups closer. …given the weight of the apparatus, both capuchin monkeys needed to pull the bars simultaneously, and high levels of cooperation were observed. To ensure that this cooperation was not merely a chance occurrence, the researchers showed that the animals performed much better when they could see each other through the partition, suggesting that each monkey understood the other’s role in the tasks and that they maintained their cooperation through forms of visual communication.

(Blueprint, p. 304)

Frans de Waal took this same idea and tried it with elephants. This time, careful controls were put in place to see if the elephants really understood that cooperation was required. The experiment was to put two elephants in different lanes and offer them a rope connected to a table. If they both pulled on the rope, the table would move food towards them until they could eat. The elephants quickly learned to cooperate. But was it possible that they just learned to pull the rope and they were unaware they were cooperating? So they released the elephants at different times. If the first elephant started pulling without the second elephant, then they’d know the cooperation was just coincidence. But the first elephant would wait for the second elephant before staring to pull. But what if the elephant was just learning something like “pull rope when near another elephant?” So de Waal cleverly tried releasing both elephants but with one of the two ropes coiled up visible but out of reach. In this case neither elephant made any attempt to pull the rope because they apparently knew it wasn’t possible to get the food without cooperation of the other pulling the rope.

(Blueprint, p. 304-305)

So monkeys and elephants – two of the most intelligent species of animals – are clearly learning concepts more complicated than simple associations and that require them to think through what seems to be simple causal explanations. i.e. I need my fellow elephant here or this isn’t going to work because it’s too heavy for me alone.

More to the point, in both of these examples, it does not seem possible for the animals to have learned the required behavior via memes since they were being put into novel controlled circumstances. So it’s unclear how behavior parsing would explain these experiments since it is based in memes.

So we do find that some animal behavior is difficult to explain without referencing a level of intelligence that includes simple causal explanations.

However, a word of caution is strongly warranted here. Consider that prior to Richard Byrne’s for on Behavioral Parsing that there were many more examples of animal memetic behavior that seemed to require us to conjecture the existence of animal explanatory knowledge. Or put another way, it wasn’t that long ago that our best explanation was that animals did have explanatory knowledge and it’s no longer so obvious that is the case.

I do not know if Byrnes’s theory of behavior parsing could also explain every example here since they seem to involve cases where no imitation is involved. My sister’s dog has no other dogs to imitate and isn’t meaningfully imitating the humans. The squirrel was believed to be in a novel circumstance that it couldn’t have previously learned from experience nor from other squirrels. The elephant and chimp experiments were controlled experiments with novel situations that required cooperation rather than imitation.

This is an area where additional study is needed to see if Byrne’s theory can cover all such observed behavior or not. I’ve recently added his research papers to my reading queue to see if I can look into this further. I’m currently hesitant to draw conclusions one way or the other.


[1] Behavioral Parsing

Here are the relevant quotes from Deutsch in The Beginning of Infinity.

If behavior is impossible to imitate without prior knowledge of the theory causing the behaviour, how is it that apes, famously, can ape? They have memes: they can learn a new way of opening a nut by watching another ape that already knows the way. How is it that apes are not confused by the infinite ambiguity of what it means to imitate?

BoI, p. 405

Deutsch goes on to say:

Apes are capable of recognizing a much larger set of possible meanings. Some of them are so complex that aping has often been misinterpreted as evidence of human-like understanding. For example, when an ape learns a new method of cracking nuts by hitting them with rocks, it does not then play the movements required to crack the nut sequence like a parrot does. The movements required to crack the nut are never the same twice: the ape has to aim the rock at the nut; it may have to chase the nut and fetch it back if it rolls away; ti has to keep hitting it until it cracks, rather than a fixed number of times; and so on. …

Such activities may seem to depend on explanation — on understanding how and why each action within the complex behavior has to fit in with the other actions in order to achieve the overall purpose. But recent discoveries have revealed how apes are able to imitate such behaviours without ever creating any explanatory knowledge. In a remarkable series of observational and theoretical studies, the evolutionary psychologist and animal-behaviour researcher Richard Byrne has shown how they achieve this by a process that he calls behaviour parsing (which is analogous to the grammatical analysis or ‘parsing’ of human speech or computer programs.)

BoI, p. 407

One Reply to “Do Animals Have Explanatory Knowledge?”

Leave a Reply

Your email address will not be published. Required fields are marked *