Live data from Hacker News

“Pain Is the Only School-Teacher”

gwern.net

51–60 of 60 posts

Re: “Pain Is the Only School-Teacher”

#51

Fools learn from experience. The wise learn from others' experience. (even knowing this, I am usually a fool. Also, could someone please tl;dr "inner vs outer losses" for me? advthanksance) Edit: am I properly interpreting "pain as grounding" to be somewhat parallel to the I term of a PID controller?

The problem with learning from others' experience is... where does this learning process come from?

Learning from watching others is actually quite difficult. As Terence said, "when two do the same, it's not the same." No set of pixel observations will be identical. How do you map a third-person view to your own first-person view? You need sophisticated algorithms before 'learning from others' experience' is even an option. We struggle to get robots to learn from imitation. Human children routinely 'over-imitate' because they can't distinguish what part of a sequence of actions is necessary and what parts are optional or can be sloppier and still achieve the goal. Indeed, how do you even know what the 'goal' was? People differ greatly in abilities, preferences, and knowledge, so you would seem to require theory of mind just to begin. (I'm reminded of a point about kittens I saw argued: cats can't learn from observations, and instead, when a kitten 'imitates' their mother, they are actually simply becoming interested in the same object or place, and then independently inventing, by the usual cat trial and error, whatever useful behavior it was.)

Re: “Pain Is the Only School-Teacher”

#52
post #50
post #24

Earlier quoted context omitted.

I'm having trouble seeing it as fallacious and what (I think?) you're describing there doesn't seem to capture what I was taught nor the usage I'm accustomed to encountering in the literature. Rather I see it as splitting along an intuitive boundary - namely a whole host of low level biological responses (seriously there's a lot) versus higher level psychological (ie perceptive, conscious, whatever) phenomena. This s…

> My understanding is of nociception being a biological phenomenon that can be measured by proxy via behavior...At absolute minimum it seems weird to me to talk about an unconscious patient experiencing pain. Surely we benefit from being able to draw a distinction here? It may be weird, but I think it's very valuable to have a taxonomy which lets you ask about kinds of pain which are not merely yoked to immediate ver…

I can't escape the impression that you aren't using the same definition I am for nociception. I otherwise agree with everything you've said as far as I can tell.

> And why ... if nociception is merely behavioral effects? ...

This is definitely not consistent with the definition I'm using. Rather, I'm using nociception to refer to the low level _physiological_ responses to damage (specifically the set of molecular pathways that are in some way organized as part of a larger systemic response to said damage).

The term pain can then be assigned to a particular qualia, leaving a few other phenomena on the levels in between the two. This makes nociception easy to speak and reason about and helps avoid some of the most confusing or apparently contradictory situations but otherwise leaves the higher level stuff (pain, motivation, various other qualia) as difficult to figure out as before.

In the case of complications related to anesthesia I'd say it's squarely in a grey area that the classification scheme I'm applying here doesn't handle as well. That's due partly to the line between physiology and psychology blurring at times, and partly to the (related) fuzziness of the term unconscious as it applies to an organism's biology.

Even when you're unconscious, a great deal of your nervous system still has to be functioning on some level in order to keep you alive. Since psychology arises from your nervous system, which is in turn made up of innumerable physiological effects, then it's not inconsistent with the definitions I'm using that we can observe nociceptive pathways resulting in changes to some of the higher level systems even if someone was unconscious while they were active.

> When Tanya decides to react to burning her hand on a stove by moving it away, is she really experiencing the exact same kind of nociception that you or I experience when we burn our hand on a stove and move it away?

I'd point out that (given the definition I've been using) we don't experience nociception but rather some higher level qualia that's downstream of it. Not knowing all that much about her case other than what appears in your article (ie pain insensitivity) I honestly have no idea. She obviously doesn't experience what any of us would describe as pain. Beyond that, how can we know what qualia someone else experiences? If we assume her to be otherwise identical to us then I suppose it would depend on exactly where in the pathway that runs from nociception to pain response her genetic abnormality manifested.

> ... despite the lack of qualia, look like learning processes about avoiding damage.

Precisely! As I'm modeling it, useful (evolutionarily) nociceptive pathways can give rise to avoidance or learning processes in any way whatsoever. All that's required is that their presence increase organism fitness!

In the case of many simpler organisms I suspect there might not be any qualia or learning whatsoever (and thus no loss function, bi-level or otherwise, on the level of an individual organism). Rather, nociceptive pathways might simply trigger some set of immediate, hard coded, higher level responses. (Consider C. elegans in particular and I think this will make a lot of sense.)

Up organism complexity a bit and you might encounter learning in the form of something resembling a simple state machine. In other words, employing nociceptive pathways to form a very basic set of associations between specific environmental conditions and some sort of danger or avoidance response. This could allow preemptively avoiding something that caused damage (and thus lowered fitness) in the past. (This might or might not be bi-level depending on what other learning pathways, if any, it interacted with.)

Under such a model, it's only when you make it up to incredibly complex organisms whose behavior is governed (to at least some extent) by higher level reasoning that you might encounter things we could identify as psychology, qualia, or pain. This is where bi-level losses seem to become particularly important and also where things rapidly become incredibly complicated to model.

That being said, it seems to me that a hardwired avoidance response (ie an outer loss) is likely only necessary for initial bootstrapping (in the case of sufficiently well educated humans) or in an organism where higher level reasoning is present (and thus driving behavior) but not sufficiently advanced. For example, it seems at least plausible that Tanya might have been fine if only there were some way to imbue her with a well developed understanding of the world. Then again, maybe not - it's quite possible that other internal processes have developed a dependence on the pain response being present, and so ripping it out without doing any other remodeling might well result in a "pain sized hole" and a nonviable organism.

Re: “Pain Is the Only School-Teacher”

#53
post #48

Earlier quoted context omitted.

100% agree. There is a very nonzero amount of "punishing wrongthink with downvotes" on HN. (Even if the post follows posting guidelines.) As long as the community continues this kind of behavior, we'll have people spamming new accounts.

Usually there's something else wrong with a downvoted comment besides just "wrongthink". Since people don't like being downvoted, or seeing something they agree with being downvoted, they tend to reach for "wrongthink" a little too readily as the explanation. Certainly there are unfairly downvoted comments, but users tend to give those corrective upvotes after a while: https://hn.algolia.com/?dateRange=all&page=0&pre…

Indeed I "correctively upvote" things that I otherwise wouldn't, but whenever I do so I'm frustrated by the rudimentary nature of pretty much all the moderation systems I've come across to date.

Re: “Pain Is the Only School-Teacher”

#54
post #51

Fools learn from experience. The wise learn from others' experience. (even knowing this, I am usually a fool. Also, could someone please tl;dr "inner vs outer losses" for me? advthanksance) Edit: am I properly interpreting "pain as grounding" to be somewhat parallel to the I term of a PID controller?

The problem with learning from others' experience is... where does this learning process come from? Learning from watching others is actually quite difficult. As Terence said, "when two do the same, it's not the same." No set of pixel observations will be identical. How do you map a third-person view to your own first-person view? You need sophisticated algorithms before 'learning from others' experience' is even an…

I currently believe the third-person view comes before the first-person view. There's a substantial evolutionary pressure to be able to predict one's predator or prey; even quite limited theories of mind come in handy there. Consciousness, however, isn't required — I'll argue that the first-person view is an accident: once we have a system that is suitable for modelling others, it makes sense that the creature we have the most substantial data stream on is ourselves, so we model that as well.

This way around might explain why our self-awareness sometimes fails to be optimal.

For how little theory of mind it takes to play antagonistic games, see Shannon's 1953 mind reading machine: https://this1that1whatever.com/miscellany/mind-reader/Shanno...

(or try to play fetch with a dog, mixing two types of throws, and see how quickly it learns your tells)

Re: “Pain Is the Only School-Teacher”

#55

I don't get his point. After all there is no need for a conscious introspection from evolutionary perspective - not even for pain. There is observable behaviour driven by events from the "outside" world. That all could easily be represented by a function of some input. No qualia needed. The truth behind this is that science has no grip on certain things. And these things are the central aspects of human experience. S…

You seem to be describing p-zombies which (IMO) don't have much overlap with present day science. Higher level reasoning would appear to increase fitness in many cases. For whatever reason, our version of it appears to involve conscious introspection.

As to the point of the article - it's about modeling human learning as a reinforcement learning problem and thus pain as one of many interacting loss functions. In particular it's about modeling how these different loss functions interact with one another to achieve a desirable outcome.

Re: “Pain Is the Only School-Teacher”

#56
post #46
post #23

Earlier quoted context omitted.

I probably didn't do a good job explaining it, but using your obstacle analogy... When you say "an object in my path caused me to go around it", you're switching the cause and effect. A more accurate way to say it would be "an obstacle on the path meant that only a handful of the population can go around it". The change that allows you to go around the obstacle was not caused by the obstacle, it would have happened r…

> However, the change is only important if the obstacle is there - that's what drives evolution (selection). I understand what's being said, it's the "selects" that triggers the problem for me. I think of evolution as a system. Inputs cause the system to act in a certain way, but it's the system responding to the inputs, not the inputs choosing to get a specific output (which is what it sounds like to me to say A sel…

> why does evolution prefer the "selects" terminology? Is that common in any other sciences?

I think this is due to set theory and statistics. Terminology such as "select x such that" is standard. In fact a choice function is commonly referred to as a selector. Since evolution is essentially a biased random walk, you can describe it as a stochastic choice function (plus some other stuff) applied to the set of entities that makes up a population.

(If "selects" bothers you, what do you make of quantum mechanical "observers"? I witnessed that one cause lots of misconceptions among university and even a few graduate students.)

Re: “Pain Is the Only School-Teacher”

#57
post #55

I don't get his point. After all there is no need for a conscious introspection from evolutionary perspective - not even for pain. There is observable behaviour driven by events from the "outside" world. That all could easily be represented by a function of some input. No qualia needed. The truth behind this is that science has no grip on certain things. And these things are the central aspects of human experience. S…

You seem to be describing p-zombies which (IMO) don't have much overlap with present day science. Higher level reasoning would appear to increase fitness in many cases. For whatever reason, our version of it appears to involve conscious introspection. As to the point of the article - it's about modeling human learning as a reinforcement learning problem and thus pain as one of many interacting loss functions. In part…

I just cannot see why the introspection is needed and to what regard it extends the functional paradigma.

For the outcome of reaction - there is no introspection needed. The difference between a pain value and qualia of pain is not clear with regard to the function of behaviour and evolution.

Qualia is not simply a state. a state can also be modelled with a function. The examples with people deactivating the "warning prothetics" are no real explanation. I think a (non conscious but complex)function or set of functions could simulate human behaviour without introspection or conscious sense.

There is, from evolutionary perspective, no necessity for it. Don't get me wrong, I appreciate the authors work. I maybe just missed where the magic begins.

Re: “Pain Is the Only School-Teacher”

#58
post #55

Earlier quoted context omitted.

You seem to be describing p-zombies which (IMO) don't have much overlap with present day science. Higher level reasoning would appear to increase fitness in many cases. For whatever reason, our version of it appears to involve conscious introspection. As to the point of the article - it's about modeling human learning as a reinforcement learning problem and thus pain as one of many interacting loss functions. In part…

I just cannot see why the introspection is needed and to what regard it extends the functional paradigma. For the outcome of reaction - there is no introspection needed. The difference between a pain value and qualia of pain is not clear with regard to the function of behaviour and evolution. Qualia is not simply a state. a state can also be modelled with a function. The examples with people deactivating the "warning…

Certainly we lack (to the best of my knowledge) a rigorous model or definition for terms such as "conscious" and "qualia". That doesn't tell us anything about whether introspection or consciousness is necessary or not, just that our understanding is lacking.

The fact that humans seem to have it doesn't say anything about it's necessity (or lack thereof). It only speaks to it's presence in our particular case.

I would challenge your assertion that "qualia is not simply a state". In the absence of a rigorous model how are you backing such a claim? Similarly, how are you justifying your assertion that human behavior could be simulated without introspection or consciousness? For that matter, how could anyone justify a claim to the contrary?

It seems to me that (at present) we simply lack the knowledge to meaningfully reason about generalizations such as introspection or consciousness being necessary or not.

As to why the author brings qualia in to it - my guess would simply be "because humans seem to have it". Sure, you could write about bi-level loss functions from a purely mathematical standpoint. The point (IMO) is to tie that mathematical model back to a concrete example in the real world in order to draw inspiration for the design of more capable RL algorithms.

It occurs to me that you might be interested in this recent essay on the topic. (https://slatestarcodex.com/2020/06/01/book-review-origin-of-...)

Re: “Pain Is the Only School-Teacher”

#59
post #7

i think one evolutionary purpose of chronic pain is to remind about the event which caused it. for example, every time it rains i'm reminded of all the events which caused fractures. hopefully that helps with not repeating them.

Sorry for the offtopicness, but could you please stop creating accounts for every few comments you post? We ban accounts that do that. This is in the site guidelines: https://news.ycombinator.com/newsguidelines.html . You needn't use your real name, of course, but for HN to be a community, users need some identity for other users to relate to. Otherwise we may as well have no usernames and no community, and that woul…

there's something about hn that i can't seem to remember my pw (also, tbi). whenever i'm on a new device, i register. sorry about that.

Re: “Pain Is the Only School-Teacher”

#60
post #58

Earlier quoted context omitted.

I just cannot see why the introspection is needed and to what regard it extends the functional paradigma. For the outcome of reaction - there is no introspection needed. The difference between a pain value and qualia of pain is not clear with regard to the function of behaviour and evolution. Qualia is not simply a state. a state can also be modelled with a function. The examples with people deactivating the "warning…

Certainly we lack (to the best of my knowledge) a rigorous model or definition for terms such as "conscious" and "qualia". That doesn't tell us anything about whether introspection or consciousness is necessary or not, just that our understanding is lacking. The fact that humans seem to have it doesn't say anything about it's necessity (or lack thereof). It only speaks to it's presence in our particular case. I would…

> Certainly we lack (to the best of my knowledge) a rigorous model or definition for terms such as "conscious" and "qualia". That doesn't tell us anything about whether introspection or consciousness is necessary or not, just that our understanding is lacking.

Yes. In (theoretical) Computer Science, one of the first things students learn are the limits of their domain and the limits of mathematics and models in general. Keywords: Halting problem

In other domains (e.g. biology, psychology etc.) they don't bother with such meta physics :) They use very primitive models (e.g. game theory) to map very complex processes. They forget the proverb that all models are wrong, but some are useful. And if that complexity gap is to big, the models just model randomness.

Long story short: I doubt it is possible to build a formal model to describe qualia or introspection.

> The fact that humans seem to have it doesn't say anything > about it's necessity (or lack thereof). > It only speaks to it's presence in our particular case. > I would challenge your assertion that "qualia is not simply a state". > In the absence of a rigorous model how are you backing such a claim?

As I said, there is no "connection", no "dependency" to evolutionary processes. It is decoupled of what we call evolution. I would rather think it is a property of matter or physics, than one of biology or even cognition.

On the other hand, take the usual examples. A color can be described as a certain wavelength. You can model that. You can build a machine, that detects / "percepts" colors and have an inner state of that "color". What is needed in addidition, to create, what we (as humans) perveive as colors? Is it a matter of complexity? Like, if a lot of processes interact and supervise each other, then there are 3rd level effects, like with magnetism and electricity. And one of this is a new emerging "force": consciousness.

> Similarly, how are you justifying your assertion that > human behavior could be simulated without introspection > or consciousness?

Maybe not in that complexity, but it's possible to simulate agents artificially in an artificial world. To simulate reality, you need a "super" reality and more energy than is existing. Modelling is costly. You have to sacrify through "abstraction". Maybe consciousness is just possible at this level of immediacy when everything interacts with everything in real time.

> For that matter, how could anyone justify a claim > to the contrary?

If it could model or emulate a certain mental state at any time of a given person and communicate that state. So that the person could verify it. Like a weather forecast. It would be a sign in the direction. But of cause a real "proof" is not possible.

Post reply on HN