This is part of a series on Simona Ginsburg and Eva Jablonka’s book: The Evolution of the Sensitive Soul, a book focused on the evolution of minimal consciousness. This particular post is on the capabilities Ginsburg and Jablonka (G&J) see as necessary to attribute consciousness to a particular species. The capability they focus on is learning, but not just any kind of learning, a type of sophisticated learning they call unlimited associative learning.
There are many different types of learning, but they can be grouped into two broad categories: non-associative learning and associative learning, with associative being the more sophisticated.
Non-associative learning includes habituation and sensitization. Habituation is when a sensory receptor responds less frequently to a constant or repetitive stimulus. It’s why you don’t feel your clothes against your skin (until I called your attention to it) or the pressure of the piece of furniture you’re sitting in against your body.
Sensitization is the opposite. If there is no stimulus for a long time, the sensory neuron is more likely to respond when one suddenly arrives. Or if it arrives in an unexpected pattern (such as the feeling of something crawling on your leg). Or if the previous stimulus was painful, then a relatively mild stimulus may still lead to an intense reaction.
Non-associative learning takes place in all kinds of living systems, including unicellular organisms. In animals, in the cases I described above, it actually takes place in the peripheral nervous system. Although it also happens in the central nervous system. More sophisticated learning is built on top of it.
Historically, associative learning has been categorized into two categories: classical or Pavlovian conditioning, and operant or instrumental conditioning.
Classical conditioning is best exemplified by the case of Ivan Pavlov’s dogs. Initially, in an experiment, the dogs would salivate when food was presented to them. But if each time the food was presented, a bell was also rung, the dogs would start to salivate on the bell ring. Eventually they would salivate on the ring so even if no food was presented. Classical conditioning is association of two sensory stimuli, the conditioned stimulus (the bell) with the unconditioned stimulus (the food).
Operant conditioning involves association between an action and a reinforcement stimulus. For example, if a rat in a cage, through random action and exploration, accidentally jumps on a lever, and a food pellet is released, the action (pressing the lever) becomes associated with a reinforcement stimulus (the food). For it to be a reinforcement, the stimulus must involve some existing value for the organism, either attractive (like food) or aversive (like electrical shock).
G&J, somewhat pushing back against the traditional nomenclature, labeling classical conditioning as “world learning”, because it involves association between external stimuli. They label operant conditioning as “self learning” because it involves associating an action by the self with reinforcement, a sensory stimulus.
- Non-associative learning
- Habituation
- Sensitization
- Associative learning
- Classical conditioning / World learning
- Operant conditioning / Self learning
G&J state that associative learning requires a brain. So although we might see non-associative learning in creatures like ctenophores (comb jellies), we only see associative learning in creatures with some sort of central coordinating system. That said, the definition of “brain” here is fairly liberal. So many worm like creatures with slightly larger ganglion toward the front of their body seem to meet the standard.
(I found this brain requirement surprising, since classical conditioning is often said to be widespread. But after reading G&J’s assertion, I tried to track down cases of classical conditioning in primitive organisms. The main example was a starfish; G&J mention the one study showing it but dismiss it for methodological reasons. They also briefly allude to studies finding it in unicellular organisms, but don’t seem to find those studies convincing.)
Primitive creatures generally only have what G&J call limited associative learning (LAL). With LAL, the associations that form are relatively simple. Although “relative” is a key word here, because even with LAL, things can get complex pretty fast.
But this isn’t the type of learning that signals minimal consciousness. For that, we need a type of learning that allows associations between compound stimuli integrated across multiple modalities (hearing, sight, smell, etc) and complex combinations of motor actions. When the capabilities for these types of associations start to arise, the possible combinations quickly increase exponentially, becoming virtually unlimited.
It is this type of learning: unlimited associative learning (UAL) that G&J see as a key indicator of minimal consciousness. UAL requires sensory integration through multiple hierarchies, forming an integrated sensorium. It also requires integration across possible motor systems, an integrated motorium. And the sensorium and motorium become integrated with each other, with a concept G&J refer to as association units. (G&J don’t use the words “sensorium” or “motorium”, but I find them helpful here to summarize a lot of detail.)
Each layer in the sensory hierarchies make predictions based on the signals from lower layers. The lower layers respond with prediction error signaling, making the communication between each layer both feed forward and feed back in a recurrent fashion. It’s with this sustained recurrent signalling that temporal thickness and synaptic plasticity is achieved, leading to memory and learning. And when it spreads to motor systems, we get the global workspace effect.
It’s important to note that G&J do not claim that UAL is minimal consciousness, only that it is a key indicator of it. In order to be capable of UAL, a species must have the underlying architecture, including the attributes listed in the last post.
However, UAL represents crucial capabilities that likely make minimal consciousness adaptive. While it’s possible to see animals that are minimally conscious who, due to injury, pathology, or immaturity, show signs of minimal consciousness but aren’t capable of UAL, the healthy mature members of the species should be capable of it. In this view, UAL is a key driver of the evolution of minimal consciousness.
In many ways, UAL resembles one of the criteria that Todd Feinberg and Jon Mallatt used for affective consciousness in their book, The Ancient Origins of Consciousness. Feinberg and Mallatt called this criteria “global non-reflexive operant learning”. (Although they didn’t elaborate on this, and I didn’t find them to necessarily be consistent with the “global” or “non-reflexive” part in the studies they cited.)
As many others do, G&J take issue with Feinberg and Mallatt dividing primary consciousness up into three separate divisions: exteroceptive, interoceptive, and affective consciousness. For G&J, there is only one consciousness, which at any time might be focused on exteroceptive, interoceptive, or affective content.
That being said, G&J reach conclusions very similar to Feinberg and Mallatt’s on which species have minimal consciousness: all vertebrates, including fish, amphibians, reptiles, mammals, and birds, as well as many arthropods such as ants and bees, and cephalopods such as octopusses.
In the last post for this series, we’ll discuss some additional areas that G&J explore, and I’ll provide my thoughts on their overall approach.
What do you think of UAL (unlimited associative learning)? Do you think it’s a valid mark of minimal consciousness?
SelfAwarePatterns 
Whatever its relationship to sentience/consciousness, unlimited associative learning is pretty damn important! Its evolutionary potential advantage is huge, and any faculties that are important to it will enjoy a big selection boost (when paired with other faculties that may be necessary).
LikeLiked by 1 person
The advantages are huge, but so is the cost, and it must have taken intense selection pressures to generate it. One of the leading theories for the explosion of development in the Cambrian was the development of the predator / prey arms race.
LikeLike
This is a good explanation and, as you might guess, I very much agree with it. I think UAL explains most of the adaptive advantage consciousness provides. The requirement that consciousness is required for UAL (do they go as far as saying that?) is often cited in support of EM field theories of consciousness – that the EM field that arises with a certain level of brain size, complexity, and structure plays a key role in synchronous firing of neurons and entraining synaptic connections associated with learning. Hopefully my copy of the book will arrive soon. Low priority, I guess, on Amazon now.
LikeLiked by 1 person
Thanks James!
I don’t think they say anywhere that UAL is impossible without consciousness, just that the conscious framework appears to evolution’s path to provide it. And it arose independently at least two times, possibly three. It’s conceivable that the common ancestor of vertebrates and arthropods was minimally conscious (although doubtful) but celphalopods definitely seem to have evolved it independently. But the authors seem open to the possibility that you could have technological UAL without consciousness. (As always, this depends on exactly how we define “consciousness”.)
On EM field theories, the authors occasionally mention “bioelectric fields”, but frustratingly never elaborate on exactly what they mean. (Or if they do, I missed it.)
I think you’ll enjoy this book, but I found it a fair amount of work to parse, and some of it, notably the discussions about proteins, I struggled with.
LikeLiked by 1 person
I would agree the ability to form rich associations is minimal characteristic of a conscious mind (assuming a definition of consciousness that goes pretty far down the chain of the animal kingdom).
FWIW, I question the difference between “classical” and “operant” — both depend on associations, which ultimately depend on some form of memory. The first skill is to recognize the link between two things. The second skill is remembering that link. (Even non-associative learning would depend on some form of local memory.)
I noticed the rejection of the starfish study and other cases that don’t fit their scheme. The boundaries might be more fuzzy then their breakdown allows. Seems like having a neural system allows training. The better that neural system, especially having a brain, the more that can be learned.
(As an aside, I think the word “unlimited” is a poor choice.)
((And where does this leave us with “deep learning” neural nets? Does what they do constitute a form of associative learning?))
LikeLiked by 1 person
I think G&J agree with you on the distinction between classical and operant, that the distinction could come from artificial experimental conditions. Although they do somewhat accept it with their alternate labels: world learning and self learning. They prefer focusing on the distinction between limited and unlimited associative learning.
On the starfish study, actually it wouldn’t really affect their thesis, since it would only grant the starfish limited associative learning. It would only mess them up if something like a starfish were found to have the unlimited variety. (Given the starfish’s limited sensory and motor abilities, this seems unlikely.) And they seemed open to the idea that unicellular organisms could have limited associative abilities, but don’t think the data is convincing yet.
I know what you mean about “unlimited”, although I can see where it comes from. I actually prefer the “global” phrase F&M used. Although both require further explanation.
I’d say neural nets do have a very limited form of associative learning, typically world learning in G&J’s nomenclature. It would take a network of neural nets able to form associations across their individual domains to approach UAL. They’d need some sort of coordinating system, something like a global workspace.
LikeLiked by 1 person
For the record, where G&J use “unlimited”, and Mike and F&M use “global”, I have been using “arbitrary”.
As for neural nets, any form of learning is associative learning, yes? You’re learning to associate this with that. The question is whether it’s limited or unlimited. Not sure why Mike marks neural net learning as “very” limited. Learning will always be limited to the inputs available. Also, I’m not sure about neural net learning as typically World learning. Is AlphaZero atypical then? And then there’s Watson. Is Watson’s learning limited or unlimited?
*
LikeLiked by 1 person
Neural nets are limited in that they can only form associations within the domain for which the net is constructed. They are also limited in the richness of those associations. For us, a single idea connects to myriad other ideas in various ways. NNs don’t have that richness.
LikeLiked by 1 person
Yet.
LikeLiked by 1 person
Wyrd answered the neural net part. But on all learning being associative, as I noted in the post, a lot of biological learning is non-associative (habituation and sensitization).
LikeLike
“…that the distinction could come from artificial experimental conditions.”
Yeah, that’s how it seemed to me. An external distinction.
“They prefer focusing on the distinction between limited and unlimited associative learning.”
That seems mostly a matter of having the right “hardware” for it.
LikeLike
It is worth noting that classical / world is sensory to sensory associations whereas operant / self is action to sensory. And operant is time sequenced, involving some degree of learned prediction. Classical usually has to be concurrent.
LikeLike
These are all external to the brain, though. (Even action is still sensory.) The brain is still linking this to that. (Ha! For once I’m on the reductive side not drawing distinctions. 😀 I do appreciate that, from an experimental point of view, the distinctions could be useful.)
Can’t classical be time sequenced? I was just reading about how dogs identify the series of actions one might make while preparing to take them for a walk (without explicitly telling the dog). They do easily pick up on a sequence of events that lead to a specific future.
LikeLike
The action-sensory aspect means the association crosses between the sensorium and the motorium. It requires more communication between the two.
(One thing I didn’t get into is that motor actions usually cause feedback to sensory processors, even in LAL systems, so they know which sensations are caused by the organism.)
I guess the time sequencing is itself another important distinction. Classical can happen without it. Operant needs it to at least a minimal extent, with increased time requiring an increasing ability to hold and relate a concept, part of the temporal thickness attribute. Put another way, operant seems to require some degree of working memory and prediction.
LikeLiked by 1 person
Okay, on those details I’ll agree there is more distinction than I thought.
The end result is the same though, yes? Ultimately a stimulus is linked to memories, even if the map is built in different ways.
LikeLiked by 1 person
A couple of questions for your expertise.
Is UAL primarily or exclusively Hebbian learning?
Do we know how Hebbian learning actually works at the neuron level? I saw some hypothesis relating to calcium ions.
LikeLike
Good question. This is actually something I hope to touch on in the next post.
I can’t claim any authoritative answer, but, at that level, I don’t think UAL involves any mechanism that LAL or non-associative learning don’t already have. The differences are in higher level architecture.
My impression is that it’s largely Hebbian, but Hebbian theory is an approximation of a lot of stuff at the molecular level, with lots of twists and nuances. G&J discuss the possibility of epigenetic engrams that influence synaptic plasticity, and speculate that those intracellular engrams might be communicated to other neurons by non-synaptic means.
Based on what I’ve read, we know a lot about how synapses work at the molecular level, but there still remain large gaps.
LikeLiked by 1 person
Epigenetic engrams sounds more exotic than EM fields.
I think explaining how learning happens is the crux of the problem. Without some mechanism to provide quick feedback on when a circuit is correct or, at least, more correct then it is hard for me to see how assemblies of neurons wire together in any reasonable time frame. Alternate explanations might work with one or two neurons and a fairly simple stimulus but when we get into larger numbers and input from multiple senses the problem becomes a lot harder.
LikeLike
I’ll be interested to see if you still have that concern after reading the book. They present several “toy” models which I think cover it. (That paper has one, but the book builds it up gradually so you’re not having to parse the whole thing at once.) In summary, subsequent sensory stimuli and affective circuits provide the feedback.
LikeLike
If their “toy” models cover it with affective circuits, why are they talking about epigenetic engrams?
LikeLike
The stated impetus is studies like this one: https://elifesciences.org/articles/03896. But honestly, it smells like a penchant for neo-Lamarckism. I don’t know if they’re at their strongest in this area. I hope to go into more detail in the next post.
BTW, “toy” is their label for a simplified idealistic model.
LikeLike
I know what a “toy” in that sense is.
Epigenetic, as I understand it, does not apply to inheritance normally, although there is something called epigenetic inheritance which is pseudo-Lamarckian but apparently real although limited. I think it has to do with alterations of gene expression during development and post birth. I saw something recently about how a lot of what we are isn’t determined directly by inherited genes but is determined by “noise” (even in the womb) that affects which genes get turned on and off.
The article you cite about the sea slugs wouldn’t seem to have much bearing since whatever memory a sea slug has wouldn’t be associated with UAL, I would think. It also wouldn’t address how Hebbian learning works with large assemblies of neurons.
It seems to me that most learning systems need some mechanism to measure how right or correct an image match, motor movement, or decision/action is. Without it there is no way to determine if the learning is on track or has reached its end. I still have trouble seeing how that works with just a bunch of neurons triggering each other.
LikeLike
Re-reading through some of their stuff, and remembering your question above about synapses, I noticed G&J cite this paper that you might find interesting: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4874022/
As I noted above, I don’t think UAL brings in anything new at the cellular level. Many of the neural mechanisms of aplysia are similar to the neural mechanisms of a creature capable of UAL.
On “just a bunch of neurons triggering each other”, it’s all in the organization. Organization is how a collection of chemical reactions work as a living organism.
LikeLike
“the neural mechanisms of aplysia are similar to the neural mechanisms of a creature capable of UAL”.
Maybe at the individual neuron level. For UAL, it seems there is qualitatively leap in the complexity because of the larger number of neurons or some other factors involved; otherwise, the sea slug would be UAL capable.
LikeLike
That link looks really interesting. I’ll need to spend some time on that.
LikeLike