Animals Can Learn: Confusion Over Specialized Definitions

Bart once asked me if I felt a correct summary of my “disagreement with Dennis Hackethal over animals” was that Dennis believed animals don’t create new knowledge in run time (during their lifetime); they, therefore, cannot error correct and cannot learn something new whereas I (Bruce) believed animals could learn and error correct.

I was shocked. I was virtually certain that Dennis didn’t believe that animals were incapable of learning (which is always a form of error correction). I’ve had conversations with Dennis about animal learning where he mentioned the nature of classical conditioning as a learning mechanism in animals and even taught me about other animal learning mechanisms. So why was Bart asking me this?

It’s Well Known that Animals Do Learn

The idea that the complexity we see in animals could all be existing knowledge in the genome would be a ridiculous notion. The genome contains about 725 Megs of memory and that is it. It’s actually mostly repetition, so it could be compressed to a mere 4 Megs. That is not even remotely close to being enough memory for an animal’s genome to store a program complicated enough to allow it to show the extremely complex adapted behavior displayed by animals. “… have you seen a machine with the autonomy of a honey bee?” asks machine learning expert Carlos E Perez.

Would animals even need brains with 100 billion neurons if their entire runtime program was a mere 725 Megs of memory? What would the rest of the storage capacity in their brains and nervous system be for?

In reality, it is well known that animals do learn. Presumably, for higher animals like mammals, this is their primary source of knowledge, not pre-wired knowledge in the genome.

Teaching a Dog to Sit

Consider even just a simple example of animal learning. You can teach a dog to sit to get a treat. If all knowledge were stored in the genome then there would be no need to train the dog. It would come pre-wired knowing that when commanded to ‘sit’ (presumably with the knowledge of how to understand the word ‘sit’ in every conceivable language!) it should lower its back legs because this action has survival value (i.e. it will receive food.)

The Efficiency of Learning

A program that had to store every single word ‘sit’ in every single language would probably be a large program. Wouldn’t it be more efficient if evolution, instead, simply wired a learning algorithm into animals so that, after they were born, they could learn which word (‘sit’) they needed to know for their particular owner?

So it should come as no surprise that biological evolution wired learning algorithms into animals very early on because of the extreme survival value learning algorithms have. That way animals can gain knowledge via the learning algorithm, allowing the animal to adapt during its own lifetime to its own specific environment.

Only dogs need to know how to sit for treats, not wolves. Why waste a wolf’s extremely limited genome storing every single possible word that might mean ‘sit’?

Classical Conditioning is a Form of Animal Learning

And this is precisely what happened. All higher animals display one or more forms of learning. The most famous of which is known as classical conditioning.  In fact, even single-celled animals have a simple learning algorithm known as Habituation where stimulus stops causing a response if the stimulus keeps occurring. This is a form of animal learning distinct from classical conditioning and it is the reason why you forget that your clothes are rubbing against your body once you get used to wearing them.

Learning and Error Correction

I think the confusion over this subject came from an interview Dennis did with Christofer. (Text version available here.) Looking back at the interview, I think the confusion came from where Dennis explains his reasons for why he feels animals are not conscious which isn’t the same thing as learning. In the interview, Dennis connected consciousness with what he calls ‘creativity’ as well as error correction. I’ve highlighted everywhere where he mentions learning, creativity, consciousness, and error correction:

So the reason I think [consciousness and being a universal explainer] are deeply connected, and that you aren’t conscious, if you’re not intelligent [i.e. a universal explainer], is because it follows from our best explanation of what consciousness is. … it seems like consciousness has to do with error correction, because when we turn inward… and we have [to] simply observe what it is that we’re conscious of… it seems to always be the thing that we’re looking to understand or that we’re looking to improve or correct. So for example, if you’re a child and you’re learning how to ride a bicycle, you will be acutely aware of all the minute little details: you have to learn how to pedal, and at the same time, you have to learn how to steer and how to keep your balance. This is a very conscious, effortful process. But as you learn, as you correct the errors in your balance in your pedaling, your awareness is drawn more and more away from this process, until you get so good at it that you don’t even have to think about it anymore.

So now when I ride my bike, I’m not aware really that I’m riding my bike, I focus my attention on the road, on the things that I need to identify so that I can correct errors. If I then run over a pothole, I become very aware of that situation. Or, like Popper said, if you go up a flight of stairs, and you reach the end of it thinking that there’s one more step, you also become very aware of the situation. So consciousness seems to have to do with disappointed expectations and the identification and correction of errors. And all of that is part of the creative process. And that is why I think that consciousness and creativity are deeply intertwined and I don’t think they’re separable neatly. So, that is why I conclude that animals that aren’t… that aren’t creative also aren’t able to be conscious and therefore suffer and so forth.

(Emphasis mine)

So Dennis conjectures that ‘creativity,’ error correction, and consciousness all go hand in hand. He uses an example of this via learning to ride a bike, but he doesn’t claim animals don’t learn.

I did find one place in the interview where perhaps Dennis seems to be claiming animals do not learn in his example of a dog that could play Jenga. He does here claim the dog did not learn anything. (A claim I refute in this post.) And in fact, animals do sometimes have pre-wired behaviors, so a specific claim to a specific behavior being in the genome is at least plausible. But nowhere do I see a general claim that animals do not learn.

However, I can see why this interview left Bart with the impression that Dennis thinks animals contain all their knowledge in their genome rather than most of it in their nervous system. Since Dennis is connecting error correction to consciousness, I can see why this might leave one with the impression that he’s claiming that only conscious beings do error correction. If this were literally true, then animal learning would be impossible because all learning involves error correction. Taken quite literally, the above passage would then be a claim that animals do not learn at all. If that were the case, then obviously all of their knowledge must come from somewhere else, and the genome is the only option remaining.

Error Correction and Consciousness Aren’t The Same Thing

In reality, error correction happens all the time with non-conscious processes. Biological evolution contains ‘error correction’ but no one believes biological evolution is conscious. It’s also well known that the immune system works via the same evolutionary (i.e. error correction) process. In fact, Campbell pointed out that all existing algorithms that create adapted information do so via this same evolutionary process which includes error correction. [1]

Consider also that humans do error correction subconsciously (unconsciously) all the time. Inexplicit knowledge is primarily gained and error-corrected entirely unconsciously. Even the very example Dennis used of riding a bike involves primarily error correction that is unconscious — as is true for all implicit knowledge.

I used the very same example of riding a bike in this post here but to make the opposite argument. Riding a bike requires constant error correction to (among other things) keep your balance, but the error correction is done unconsciously once you’ve acquired the skill. So error correction does not require consciousness. It’s rather the other way around, consciousness requires error correction — as do all evolutionary processes.

So this is the first problem with Dennis’ statement that is leading to confusion. He connects error correction with consciousness and leaves the impression that only conscious beings can error correct — and thus animals (which he believes aren’t conscious) can’t.

What is “Creativity”?

Another source of confusion from this interview comes from how Dennis uses the word ‘creativity’ in a specialized way. Consider this quote where he argues that some dog’s don’t need to learn to what pointing is (backed up by a study):

So, at that point, well, the knowledge is again provided by the genes and no creativity is required on the part of the dog. So, I wouldn’t consider… it comes back to the sophisticated knowledge thing, intelligence versus smarts. The dog is smart because it can understand pointing, it’s just… I think it’s genetically driven.

Recognizing pointing may well be knowledge evolved into the genome of some dogs, but Dennis goes on to claim that animals have no ‘intelligence’ at all but only ‘smarts’ which he (in the quote above) previously connected to knowledge in the genome (at least in that particular case.) So this comes across like he’s saying animals only utilize knowledge in the genome. But I don’t believe this was the point Dennis was actually intending to make.

I believe the issue here is the special way in which Dennis uses the term “creativity” differently than how the word is commonly used. Dennis seems to be utilizing the term similar to how David Deutsch does in The Beginning of Infinity. Deutsch explicitly defines ‘creativity’ in a special way:

Creativity: The capacity to create new explanations (p. 30)

However, this isn’t how the word ‘creativity’ normally is used. The word has been around for a long time, long before Deutsch choose to define it that way.

Consider that most artistic skill is inexplicit and thus has no ‘explanations’ at all. Deutch’s specialized definition for ‘creativity’ would make most art ‘non-creative’ because it’s not about creating new explanations but instead primarily about developing inexplicit skills. Likewise, when Lee Sedol beat Alpha Go using “the God Move” — a creative new move never previously seen — he has no idea how to explain how he came up with it. It simply popped up out of his unconscious mind as the best move. This knowledge Sedol created (which changed the way people play Go!) wouldn’t count as ‘creativity’ under the way Deutsch defines ‘creativity’ because it was inexplicit knowledge instead of explanatory knowledge.

For that matter, Alpha Go itself was quite creative — coming up with moves and strategies no human being had ever seen before and changed the way people play Go — yet its now-famous learning algorithm creates no explanatory knowledge at all.

Even Deutsch himself does not consistently use the word ‘creativity’ to refer solely to ‘creating new explanations.’ As I pointed out in footnote one of this post, he refers to biological evolution as ‘creative’ despite it having no explanatory knowledge either. So it’s a mistake to simply assume that the word ‘creativity’ always only refers to the human ability to create explanations.

Likely Dennis was trying to say that the dogs don’t require any learning that required explanatory knowledge to play Jenga (or to understanding pointing.) But it must have learned via classical conditioning (and other means) how to balance the blocks such that the pile didn’t fall down — not unlike how a child learns to balance a bike. This is a form of animal learning, of course, but includes no explanatory knowledge.

The Perils of Making Up Specialized Definitions

I’ve taken exception in past posts to insisting that all definitions for a word but your own are ‘wrong.’ That is the epistemological error of Word Essentialism. (See here for several examples of me arguing against this practice.)

However, it’s not necessarily wrong to use an existing word in a specialized way like Deutsch (and I believe Dennis) is doing, so long as you don’t then make the mistake of thinking that is the only possible usage of the word.

Scientists explore new and innovative ideas and they need to use words to point to those ideas. Often there is no word that means only the idea they are trying to express and has no other possible meanings. So scientists face a difficult problem: do they make up an entirely new word? (Like Charles does in this post.) Or do they pick the closest word they can find and then specify that they are using that word in a specialized way? Both approaches are likely to lead to confusion, so it’s a tough choice.

But the end result of re-using an existing word in a new creative way is that some confusion tends to result because Popper’s epistemological rule of word nominalism has now been broken. People won’t be able to help but intermix the more common understanding of the term from the special ‘marked’ (as it’s called in the literature) version of the word the scientist is now using. The end result is confusing the two concepts and accidentally intermixing the two meanings.

To put it concretely: animals do learn, thus they do error correction. Learning is always a form of knowledge creation, even for animals. The word ‘creativity,’ when used in its regular non-specialized way, often refers to many types of knowledge creation that have nothing to do with the uniquely human ability to create explanatory knowledge. Thus animals are creative in the regular common usage of the term because they utilize learning and error correction to create knowledge.

But Deutsch did not mean the common usage when he uses the term ‘creativity’. He means explicitly the creation of new explanations – an ability that animals presumably lack. Thus under Deutsch’s specialized definition of ‘creativity’ animals are not creative and Dennis’ statement isn’t incorrect under that more narrow specialized understanding of the word ‘creative.’

It’s tempting to ask ‘Well which is it? Are animals creative or aren’t they?’ But then you’re being a Word Essentialist. Both usages of the term are ‘correct’ within the context in which they were originally used. Deutsch explicitly defined ‘creativity’ to be more narrow than the usual term before starting to use it that way throughout his book. The mistake is only to think that every other use of the term must follow this usage.

What is “Knowledge”?

Adding to the confusion is that Deutsch is also using a specialized definition of the word ‘knowledge.’ Most people probably think of ‘knowledge’ as being a synonym for ‘adapted information.’ Animal learning is adapted information. Machine Learning also creates adapted information. Both of these are therefore examples of ‘knowledge creation’ that are non-explanatory. [1]

But note that Deutsch’s also uses the word ‘knowledge’ in a specialized sense:

Moreover, these two types of information [human brains and DNA] that they respectively evolved to store have a property of cosmic significance in common: once they are physically in a suitable environment, they tend to cause themselves to remain so. Such information – which I call knowledge

BoI, p. 78

Deutsch specifies “which I call knowledge” making it clear that he is utilizing the term in a specialized way different than how it’s normally used. Obviously, the word ‘knowledge’ pre-existed Deutsch and had a meaning prior to him using it in this specialized way.

So Deutsch’s specialized definition of ‘knowledge’ covers only forms of knowledge that tend to keep themselves instantiated — such as DNA and human memes. But there are other forms of ‘knowledge’ that don’t fit that specialized definition.

Conclusion: So Much Confusion For a Word

Once we’re using a single word to point to two (or more) different concepts, it’s easy to intermix the two meanings. If you have in mind Deutsch’s view that animals are “not creative” (i.e. don’t form new explanation), you may accidentally fall into the error of thinking animals create no knowledge and thus don’t learn.

But if animals don’t ‘learn’ or ‘create knowledge’ this leads to the fallacy that an animal’s entire knowledge must be in its genome. This seems to be how Bart came to believe Dennis thought all animal knowledge was contained in the genome.

In reality, animals have an incredible ability to learn that far surpasses anything we currently know how to do using our best algorithmic learning techniques. I documented this at length in this post as part of my series on Reinforcement Learning. (Part 1 found here.) Animals are capable of error correction. But this doesn’t tell us anything about if animals are or aren’t “conscious” (or even if they feel qualia — which may not be the same thing as “consciousness”) because error correction doesn’t imply consciousness anyhow.

Footnotes:

[1] Campbell goes one further and explains why it would be impossible to have evolved behaviors in the genome without first evolving an ability to learn.

[2] Machine Learning is also adapted information.

Incidentally, I have a strong suspicion this is how to resolve the apparent disagreement between Campbell and Popper and Deutsch over if knowledge creation is rare or ubiquitous. I believe Campbell and Popper used the word ‘knowledge’ in the regular sense which means something like ‘adapted information that solves a problem.’ This kind of knowledge creation is ubiquitous just like Popper and Campbell believe. But Deutsch was (possibly) using his specialized definition, which is rarer. That might be why he wasn’t so sure Artificial Evolution algorithms created ‘knowledge’ if he only meant it in this more narrow sense.

This is something I wish to explore in more detail in a separate post as it’s not straightforward if that is the case or not. In all honesty, it’s hard to make sense of what Deutsch actually meant. But in any case, machine learning and animal learning are cases of knowledge creation just like Campbell and Popper believed if you are using their understanding of those terms.

2 Replies to “Animals Can Learn: Confusion Over Specialized Definitions”

Leave a Reply

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