Happy Monday all.
I feel like a chapter ends whenever we pass the ‘official’ deadline for the end of marking, even if we’ve got 25% of papers to get across the finish line and get marks on by the middle of August.
Once that official deadline falls, I’m so conditioned to bells and whistles that it feels like summer starts. 30 years of exam marking will do that to a person. It doesn’t matter that I’m about to get into the bunfight that ensues as we all try and do our favourite questions that pay the best, quitting whenever we feel like it. Plus, I’ll be marking long after that, I imagine, and review marking starts on the 21st August and will fill my days through to mid-October.
So you can imagine my kind of ‘school’s out for the summer’ moment when the clock rolled over on Friday into the free-for-all instead of the obligations. I feel kind of free and liberated, like I don’t have a to-do list that covers my living room wall.
This also means that, instead of trying to squeeze in a fear-writing session a week between my other stuff, I can now start trying to squash in two or even three.
The madness.
Yes, I know my social posts have expired and gone extinct. Yes, I know I’ve only got a few more Thursday substacks sitting in the wings for those of you who support me. Yes, I know I’ve still got five modules to finish off for the fear course, ideally before people have finished the modules I’ve already hit publish on that are good to go. And yes, I know I’ll be relaunching the Frustration Masterclass in six weeks’ time or so.
I’m Emma. I have pipe dreams that make me cry every time they escape my grasp.
Anyway, I’ve been writing some of the more cognitive elements of the fear course which is right up my alley. Appraisals, salience, discrimination, feature identification, illusions, sensory integration, decision-making, stored information, memory, trauma - you know this is all up my street.
And a little one on prediction.
That’s what I’ve been focusing on this week.
Like quite a bit of the fear course, I wouldn’t have bothered to have gone more than ten minutes on prediction if it hadn’t been for the utter bullshit amount of misplaced knowledge that circulates in the dog world. Also, I wouldn’t have bothered, because there’s a lot of theoretical physics and maths. Like a lot. It’s all stuff I know, but not always confidently enough to explain.
It’s also highly, highly theoretical since the stuff that dog trainers tend to take comes from predictive machine learning, not predictive learning in biological organisms.
I don’t know why dog trainers would decide predictive machine learning - the domain of people like Geoffrey Hinton - the so-called Godfather of AI - and Peter Dayan - is not only applicable to dogs, but also a factual explanation of how dogs learn.
It isn’t.
There are simple parallels, but prediction learning and reinforcement learning in machines are not explanations of how dogs learn.
I don’t know why I’d have to say such a thing outloud.
Well, I do.
Because there’s a bunch of dog trainers harping about reinforcement contingency gaps and prediction errors and this makes me want to poke my eyes out with cocktail sticks, because there are SO many reasons these are NOT models of how animals learn. I can’t even tell if the trainers who are doing this are a) aware that these are models of artificial learning in machines and programming or are b) cynically relying on the fact that other people don’t know in order to make themselves sound knowledgeable. Or both, I guess. Or c) think dogs are machines. Given the very, very large proportion of the trainers talking about prediction errors and reinforcement contingengy gaps calling their dogs ‘K9s’, I’m not really entirely sure they don’t think their dog is a machine.
I also know how deeply theoretical prediction learning and (AI) reinforcement learning is. Much of it marries up physics, maths and computational models into simulations that, while make a lot of sense about how organisms organise their behaviour, are theoretical. Or used to shape learning in AI models. Theoretical.
And hard.
The go-to guy on this in organic, biological learning is Professor Karl Friston and his free energy principle. This theory explains how the brain is a prediction-making organ.
It’s just about sensible to me, although I did have to get some maths friends to explain the aspects of Bayesian prediction, Shannon’s information processing and a bunch of other bits. I did the philosophical bits myself, thanks for asking. But I had to on-board some proper geeks for the theoretical physics bits, the explanation of Markov blankets, Markov chains and a bunch of other probability stuff. Proper maths boffs. And one of those maths friends is a senior actuarial analyst with a first in Maths from UMIST and a prize for outstanding results in his final exams. He only got a first because he was on a lot of drugs at the time. His whole job is literally probabilities. He also goes skiing a lot. I don’t know how anyone who deals in probabilities can take up such risky sports, especially as he’s now getting on a bit and should know better. I asked him to explain these bits of Friston’s theory to me, and he did a fair enough job so that more of it made sense than it did. I managed the biology & neuroscience bits myself. Phew.
So you can see part of the problem on predictions even when applied to biological organisms like humans and like dogs. It’s hard. It’s niche philosophical & theoretical neuroscience. It’s very, very maths. Even if you are a wunderkind for maths and you won prizes even if you were going to a lot of raves at the time.
That could be scary for some. Me, for one.
Clearly NOT scary to the kind of dog trainers who try and do eye-rollingly silly clips on TikTok or Instagram about the importance of prediction errors for learning who don’t seem to realise they’re talking about the AI version, not the organic beings version. But potentially scary for anyone with half a brain cell.
Professor Matthew Cobb in The Idea of the Brain said that ‘Despite the attraction of Friston’s approach for the mathematically minded – I happily admit it is beyond my grasp’. I like Matthew Cobb. I love this quote. When smart neuroscientists and biologists say that things are beyond their grasp, it gives the rest of us hope.
We could do with more people saying that.
Cobb also raises a point that ‘a fundamental problem remains’ with Friston’s theory which is that data is almost non-existent in support of it.
Oh well.
Don’t let a lack of data stand in your way!
That certainly doesn’t stop any dog trainers harping on about prediction errors and machine versions of reinforcement learning. I mean I get the confusion. It’s very irritating that our Skinnerian cousins in AI borrowed from similar aspects of operant learning; it’s even more irritating they call their flavour of learning ‘reinforcement learning’ which could bamboozle a person into thinking they’re the same. And they are cousins, in a similar kind of way to Orthodox Christianity being a cousin to being a Quaker, I guess.
But that leaves us wondering what Friston’s approach to prediction and the brain is – and that of other thinkers and researchers who are influenced by it – as well as what it suggests and why it’s so popular in certain quarters of the dog world, even if they don’t even know it comes from Friston, Dayan, Hinton, let alone the complexities of the free energy principle.
Why would such theories end up diluted beyond all recognition in the words of a man-with-camouflage-and-malinois on TikTok talking about why your dog needs to fail?
It also leaves us with the very big, bad question of what all that even means for our fearful and anxious dogs.
Cobb later goes on to say in his section about Friston’s work and how the brain is a prediction-making organ that ‘something like’ Friston’s theory ‘seems certain’ even though ‘the theoretical generalisation of this assumption to explain the whole of the brain remains speculative’.
Quite.
It’s probably a really, really good theory even if there’s no real evidence of it. It’s why I like it. It is a very good theory. And Friston is surprisingly easy to listen to about it.
Good theories without evidence are not new in psychology either. Not just neurocognitive philosophy. We’re all pretty happy with talking about consciousness, and evidence for that is thin on the ground. Lots of psychology involves a bit of a leap of faith.
So I’m not cross about dog trainers diluting it for a 2-minute video.
I’m cross because while my maths friend was trying to explain theoretical maths to me, I was busy realising that the quarter from which these bad references to prediction errors were coming was … those people who like to use shock, prong collars and the likes, and call their dogs K9s unironically.
There’s one group of dog trainers who like to talk about error and prediction stuff. And all of them have access to violent weapons with which to harm their learners who they enjoy treating as emotionless robots.
Not only that, these ‘men with beards’ or ‘men in camouflage’ or silly ponytails professing to be alphas in their relationship are frequently buoyed up by two women who have PhDs in loosely related science subjects who are also putting out content related to what they’ve termed the brain’s ‘need’ for error in authoritative and very, very wrong ways.
What it seems they’ve taken from Friston (if they’ve even heard of the man) or Dayan & Hinton is the simplest, most reductionist bits. And they’ve missed the main point - that our predictions are shaped by what’s happened in the past, because behavioural choices are embedded in a series of events including the behavioural oddities our ancestors had which were passed on through genes and traits, through developmental processes, through experience and learning.
No choice we make about what we’re going to do can be removed from the situation & body in which it occurs.
I’ve noticed it’s the same with game theory by the way. Prediction errors and game theory tend to be used most often by dog trainers who seem to be trying to sound smart and sciency in order to justify treating dogs as if they’re machines learning to play roulette better and optimise outcomes in probability learning ways that would get them banned from Vegas.
So while my friend was helping me really get my head round the Bayesian bits of Friston’s theory, I was doing the stuff I’m good at - namely finding where stuff came from on the internet, creating timelines of popularity and tracking ideas. I’m like a bloodhound when it comes to tracing where ideas came from.
That’s the bit that landed me with the awfully big ‘Oh!’ moment that I was trying to simplify the free energy principle and the Bayesian brain enough to explain how that - theoretically of course! - might relate to why dogs keep choosing to bark at scary stuff (thanks, also, to my friend who sent me videos of her dog alarm barking at practically everything that has changed because of the season out on their walk, including barking at a fallen plum from a tree - I live for those videos!).
Oh, I thought.
I see.
I get it. When our model of the world is confirmed, there is no learning required.
Tick the box that says ‘Well, that happened again! Exactly like I thought it would!’.
And tick the box that says, ‘I wonder why that keeps happening to me?’ without really stopping to ask a therapist (or even an actuary!) why.
When our model of the world is out of sync with what actually happens - there is error. The prediction error, in fact. A mismatch between what was expected and what occurred.
That’s (partially) when learning is useful.
The problem is that errors need not be bad. When we think of errors, we might be tempted to think of that awful UH-URRR noise on Family Fortunes. Blue screens of death. Alarms. Sirens going off. Being wrong.
In everyday talk, errors and corrections are coded. They’re not neutral terms. And when those aversive trainers using the need to make an error as the reason to set our dogs up to fail, to shock them, to teach them that they’re wrong, they want the kind of errors and corrections that engender fear, that cause anxiety, that cause frustration.
When they talk about the need for prediction errors for learning, when they scoff at our attempts to create errorless learning environments or distress-free learning for our dogs, what they mean is that they’ve misunderstood predictions and errors and just about everything to do with it.
In trying to justify their use of corrections as necessary or required for a dog to learn, and in trying to criticise trainers like me who often work with dogs for whom error is challenging, frustrating or upsetting, they’ve not only simplified very hard theories past their usefulness, not only diluted them so far as to be the homeopathic version of maths, they’ve actually misunderstood whole notions.
Errors simply mean a mismatch between what you were expecting and what you got.
I expected to get a pay packet that matched last year’s. I got one that made me say, ‘Oh! Hello, new Magimix!’
It also - though it is patently not holiday or overtime - had a surprising sum for holiday pay which also made me say, ‘Oh! Nice! Someone’s forgot to knock the holiday pay button off. Bonus for me!’
Joy! Not euphoria, but getting an extra £50 you weren’t expecting is not to be sniffed at.
Back to that post from a couple of weeks back about burgers in bushes… that would also cause a prediction error. For Lidy, would she expect to find a burger in a bush?
No, dear reader. No, she would not.
Prediction error.
Update the model.
This is the stuff of prediction error learning that is not only equal to, but surpasses that of aversives. When Heston got a zap from an electrified boar fence back in the day, he steered clear about three days until he got another zap. Then I learned from it and walked us other ways until it was switched off. When Heston found a rotting, decapitated pigeon underneath a tree and managed to roll in it before I got to him, you can bet he made a beeline to that tree for the next eighteen months of his life.
In fact, I talked about these surprising moments being critical in our relationships with our dogs.
Those ‘oh!’ moments can be powerful things.
But it’s the emotion that makes them so.
That’s why predictive learning in actual organic matter with emotions is worlds apart from the drivel the K9 dog trainers are trying to suggest as a model of learning.
Emotions. Powerful things.
And when emotions are powerful with goodness, so much the better.
When they’re congruent with our needs, meet our goals in life, offer us potent reward, all is good.
Life doesn’t need to be filled with corrections for us to learn.
That idea is patently wrong. Being bad-wrong (in the ‘got zapped by a fence’ way) is not the only way to learn; being good-wrong (in the ‘finding stinks to roll in’ way) is also a way to learn. Neither of these are the only ways to learn. Neither are critical or crucial for the learning process, as such dog trainers claim. And if there’s an option between being bad-wrong and good-wrong, intensity matters, but reward will trump most but the ugliest of bad-wrong experiences.
I mean I’ve got a dog snoring behind me who caught and shook a cat back in 2016 and who has now spent the next nine years in pursuit of that one, heady, powerful reward.
So surprising, so goal-congruent, so in line with millennia of predatory ancestors, so delicious was that moment that all attempts by her former owner to punish her for her delightful surprise fell on stony and unproductive ground.
Oh, actually, they were not entirely stony and unproductive. That’s most likely where she picked up a bunch of redirected frustration from and would happily bite the human holding the lead for months every time she saw a cat or walked past cat smells.
I think it all just made me realise how much science is abused by those who’d also harm dogs.
I’m sure Karl Friston has no idea that his over-diluted theories have been misconstrued and misrepresented as being factual by a bunch of blokes having a masculinity crisis on social media, their ‘K9’ tools and their female sock puppets.
In fact, I’m sure that the dog trainers abusing Friston’s work in such a way have no idea it’s his, having probably been diluted through TED talks and ninety-nine degrees of Manosphere podcasts or AI-generated pap through to the kind of echo chamber in which they live where maths gets misused as an excuse to be violent and shitty to dogs. If you asked them where their notions about prediction errors came from and their so-called crucial status in learning, I have no doubt they wouldn’t know.
I’m that certain. I’d be shocked.
Now that would be a prediction error and a half.
I might even raise an eyebrow to find out they knew Friston’s name.
What I did learn is that all the chat about prediction errors & learning in the dog world, and all the wrongness I felt obliged to put a little straighter as best I could, well all that chat seems to be coming from people trying to bamboozle others and give the impression they KNOW things.
And then I felt a bit silly for getting so bogged down in drama created by others to try to get others to respect them. Urgh. I hate it when that happens.
But then I remembered, a bit like when people talk about Yerkes-Dodson, it’s one of those great litmus tests that help us understand if a seemingly scientific person is actually just a sock puppet for a bloke with a ponytail trying to sell shock collars, or whether they’re so far out of their academic lane that they might as well be on another continent.
Anyway, it all reminded me how important it is to engage the brain and to chew stuff over with friends you’ve not heard from for a while, if only to ask them how their skiing is going and whether they could explain how Bayesian probabilities help us understand how the brain might - theoretically speaking - make predictions about whether or not there’s a burger in a bush.
It also reminded me that there’s a heck of a lot of abuse of the old ‘expert power’ out there, of people who love to sound smart and right in their pursuit of creating fearful and cowed dogs, and that anybody who thinks science tells them anything for sure needs to listen to an actual scientist saying, ‘well… I think you’ll find it’s a bit more complicated than that.’
In fact, that was one reason that it was very, very good to chat to a senior actuary who skiis a lot. Just because you know the probability of injuring yourself heinously on the slopes, just because you know how silly, reckless & dangerous it is per hour skiied, it doesn’t stop you being an adrenaline junkie. You can know those probabilities intimately. You can know all the best ways to keep yourself healthy and alive, and all the best ways to die or injure yourself before your time, and you can still decide it’s sensible to jump out of planes when you’re fifty two years of age.
Actuaries, of all people, know that people are improbable, complex creatures influenced by all the emotional yums in life. Not just mathematical outcomes.
And even for those of us who also get that, trying to work out how any mathematical probability might apply to dogs who bark at fallen plums - well, that’s the fun bit.
PS in short, dogs make appraisals about the world based on their stored knowledge (including that of all their dog & wolf ancestors, and further back still via genes) and if it doesn’t match with what you thought should happen, bark at it if you’re of an anxious disposition, or gobble it up in your mouth if you’re not. If you can’t discern from that information-gathering exercise whether it’s edible or a threat, probably piss on it and walk away.
Thanks for coming to my TED Talk on why dogs bark at fallen plums and why bad dog trainers do bad things with science. It’s been great having you. Have a safe journey home.
PPS normal services will be resumed soon.