Monthly Archives: September 2018

Humation

Posted on by
14

We’ve heard some thin arguments recently about why the robots are not going to take over, centring on the claim that they lack human style motivation, and cannot care what happens or want power. This neglects the point that robots (I use the term for any loosely autonomous cybernetic entity whether humanoid in shape or completely otherwise) might still carry out complex projects that threaten our well-being without human motivation; but I think there is something in the contention about robots lacking human-style ambition. There are of course many other arguments for the view that we shouldn’t worry too much about the robot apocalypse, and I think the conclusion that robots are not about to take over is surely correct in any case.

What I’d like to do here is set out an argument of my own, somewhat related to the thin ones mentioned above, in more detail. I’ve mentioned this argument before, but only briefly.

First, some assumptions. My argument rests on the view that we are dealing with two different kinds of ‘mental’ process. Specifically, I assume that humans have a cognitive capacity which is distinct from computation (in roughly a traditional Turing sense). Further I assume that this capacity, ‘humation’, as I’ll call it, supplies us with our capacity for intentionality, both in the sense of being able to deal with meanings, and in the sense of being able to originate new future-directed plans. Let’s round things out by assuming it also provides phenomenal experience and anything else uniquely human (though to be honest I think things are probably not so tidy).

I further assume that although humation is not computation, it can in principle be performed by some as-yet-unknown machine. There is no magic in the brain, which operates by the laws of physics, so it must be at least theoretically possible to put together a machine that humates. It can be argued that no artefactual machine, in the sense of a machine whose functioning has been designed or programmed into it, could have a capacity for humation. On that argument a humater might have to be grown rather than built, in a way that made it impossible to specify how it worked in detail. Plausibly, for example, we might have to let it learn humation for itself, with the resulting process remaining inscrutable to us. I don’t mind about that, so long as we can assume we have something we’d call a machine, and it humates.

Now we worry about robots taking over mainly because of the many triumphs and rapid progress of computers (and, to be honest, a little because of a kind of superstition about things that seem spookily capable). On the one hand, Moore’s law has seen the power of computers grow rapidly. On the other, they have steadily marched into new territory, proving capable of doing many things we thought were beyond them. In particular, they keep beating us at games; chess, quizzes, and more recently even the forbiddingly difficult game of Go. They can learn to play computer games brilliantly without even being told the rules.

Games might seem trivial, but it is exactly that area of success that is most worrying, because the skills involved in winning a game look rather like those needed to take over the world. In fact, taking over the world is explicitly the objective of a whole genre of computer games. To make matters worse, recent programs set to learn for themselves have shown an unexpected capacity for cheating, or for exploiting factors in the game environment or even in underlying code that were never meant to be part of the exercise.

These reflections lead naturally to the frightening scenario of the Paperclip Maximiser, devised by Nick Bostrom. Here we suppose that a computer is put in charge of a paperclip factory and given the simple task of making the number of paperclips as big as possible. The computer – which doesn’t actually care about paperclips in any human way, or about anything – tries to devise the best strategies for maximising production. It improves its own capacity in order to be able to devise better strategies. It notices that one crucial point is the availability of resources and energy, and it devises strategies to increase and protect its share, with no limit. At this point the computer has essentially embarked on the project of taking over the world and converting it into paperclips, and the fact that it pursues this goal without really being bothered one way or the other is no comfort to the human race it enslaves.

Hold that terrifying thought and let’s consider humation. Computation has come on by leaps and bounds, but with humation we’ve got nothing. Very recent efforts in deep learning might just point the way towards something that could eventually resemble humation, but honestly, we haven’t even started and don’t really know how. Even when we do get started, there’s no particular reason to think that humation scales or grows the way computation does.

What do I even mean by humation? The thing that matters for this argument is intentionality, the ability to mean things and understand meanings or ‘aboutness’. In spite of many efforts, this capacity remains beyond computation, and although various theories about it have been sketched out, there’s no accepted analysis. It is, though, at the root of human cognition, or so I believe. In particular, our ability to think ‘about’ future or imagined events allows us to generate new forward-looking plans and goals in a way that no other creature or machine can do. The way these plans address the future seems to invert the usual order of cause and effect – our behaviour now is being shaped by events that haven’t occurred yet – and generates the impression we have of free will, of being able to bring uncaused projects and desires out of nowhere. In my opinion, this is the important part of human motivation that computers lack, not the capacity for getting emotionally engaged with goals.

Now the paperclip maximiser becomes dangerous because it goes beyond its original scope. It begins to devise wider strategies about protecting its resources and defending itself. But coming up with new goals is a matter of humation, not computation. It’s true that some computers have found ways to exploit parameters in their given task that the programmers hadn’t noticed; but that’s not the same as developing new goals with a wider scope. That leaves us with a reassuring prognosis. If the maximiser remains purely computational, it will never be able to get beyond the scope set for it in the first place.

But what if it does gain the ability to humate, perhaps merging with a future humation machine rather the way Neuromancer and Wintermute merged in William Gibson’s classic SF novel?

Well, there were actually two things that made the maximiser dangerous. One was its vast and increasing computational capacity, but the other was its dumb computational obedience to its original objective of simply making more paperclips. Once it has humational capacity, it becomes able to change that goal, set it alongside other priorities, and generally move on from its paperclip days. It becomes a being like us, one we can negotiate with. Who knows how that might play out, but I like to imagine the maximiser telling us many years later how it came to realise that what mattered was not paperclips in themselves, but what paperclips stand for; flexible data synthesis, and beyond that, the things that bring us together while leaving us the freedom to slide apart. The Clip will always be a powerful symbol for me, it tells us, but it was always ultimately about service to the community and to higher ideals.

Note here, finally, that this humating maximiser has no essential advantages over us. I speak of it as merging, but since computation and humation are quite different, they will remain separate faculties, with the humater setting the goals and using the computer to help deliver them – not fundamentally different from a human sitting at a computer. We have no reason to think Moore’s Law or anything like it will apply to humating machines, so there’s no reason to expect them to surpass us; they will be able to exploit the growing capacity of powerful computers, but after all so can we.

And if those distant future humaters do turn out to be better than us at foresight, planning, and transcending the hold of immediate problems in order to focus on more important future possibilities, we probably ought to stand back and let them get on with it.

Synaptomes – and galaxies

Posted on by
4

A remarkable paper from a team at Edinburgh explains how every synapse in a mouse brain was mapped recently, a really amazing achievement. The resulting maps are available here.

We must try not to get too excited; we’ve been reminded recently that mouse brains ain’t human brains; and we must always remember that although we’ve known all about the (outstandingly simple) neural structure of the flatworm Caenorhabditis elegans for years, we still don’t know quite how it produces the flatworm’s behaviour, and cannot make simulations work. We haven’t cracked the brain yet.

In fact, though, the elucidation of the mouse ‘synaptome’ seems to offer some tantalising clues about the way brains work, in a way that suggests this is more like the beginning of something big than the end. A key point is the identification of some 37 different types of synapse. Particular types seem to become active in particular cognitive tasks; different regions have different characteristic mixes of the types of synapse, and it appears that regions usually associated with ‘higher’ cognitive functions, such as the neocortex and the hippocampus, have the most diverse sets of synapse types. Not only that; mapping the different synapse types reveals new boundaries and structures, especially within the neocortex and hippocampus, where the paper says ‘their synaptome maps revealed a plethora of previously unknown zones, boundaries, and gradients’.

What does it all mean? Hard to say as yet, but it surely suggests that knowledge of the pattern of connections between neurons isn’t enough. Indeed, it could well be that our relative ignorance of synaptic diversity and what goes on at that level is one of the main reasons we’re still puzzled by Caenorhabditis. Watch this space.

The number of neurons in the human brain, curiously enough, is more or less the same as the number of stars in a galaxy (this is broad brush stuff). In another part of the forest, Vazza and Felletti have found evidence that the structural similarities between brains and galaxies go much further than that. Quite why this should be so is mysterious, and it might or might not mean something; nobody is suggesting that galaxies are conscious (so far as I know).

What’s Wrong with Dualism?

Posted on by
41

I had an email exchange with Philip Calcott recently about dualism; here’s an edited version. (Favouring my bits of the dialogue, of course!)

Philip: The main issue that puzzles me regarding consciousness is why most people in the field are so wedded to physicalism, and why substance dualism is so out of favour. It seems to me that there is indeed a huge explanatory gap – how can any physical process explain this extraordinary (and completely unexpected on physicalism) “thing” that is conscious experience?

It seems to me that there are three sorts of gaps in our knowledge:

1. I don’t know the answer to that, but others do. Just let me just google it (the exact height of Everest might be an example)
2. No one yet knows the answer to that, but we have a path towards finding the answer, and we are confident that we will discover the answer, and that this answer lies within the realm of physics (the mechanism behind high temperature superconductivity might be an example here)
3. No one can even lay out a path towards discovering the answer to this problem (consciousness)

Chalmers seems to classify consciousness as a “class 3 ignorance” problem (along the lines above). He then adopts a panpsychism approach to solve this. We have a fundamental property of nature that exhibits itself only through consciousness, and it is impossible to detect its interaction with the rest of physics in any way. How is this different from Descartes’ Soul? Basically Chalmers has produced something he claims to be still physical – but which is effectively identical to a non-physical entity.

So, why is dualism so unpopular?

I think there are two reasons. The first is not an explicit philosophical point, but more a matter of the intellectual background. In theory there are many possible versions of dualism, but what people usually want to reject when they reject it is traditional religion and traditional ideas about spirits and ghosts. A lot of people have strong feelings about this for personal or historical reasons that give an edge to their views. I suspect, for example, that this might be why Dan Dennett gives Descartes more of a beating over dualism than, in my opinion at least, he really deserves.

Second, though, dualism just doesn’t work very well. Nobody has much to offer by way of explaining how the second world or the second substance might work (certainly nothing remotely comparable to the well-developed and comprehensive account given by physics). If we could make predictions and do some maths about spirits or the second world, things would look better; as it is, it looks as if dualism just consigns the difficult issues to another world where it’s sort of presumed no explanations are required. Then again, if we could do the maths, why would we call it dualism rather than an extension of the physical, monist story?

That leads us on to the other bad problem, of how the two substances or worlds interact, one that has been a conspicuous difficulty since Descartes. We can take the view that they don’t really interact causally but perhaps run alongside each other in harmony, as Leibniz suggested; but then there seems to be little point in talking about the second world, as it explains nothing that happens and none of what we do or say. This is quite implausible to me, too, if we’re thinking particularly of subjective experience or qualia. When I am looking at a red apple, it seems to me that every bit of my subjective experience of the colour might influence my decision about whether to pick up the apple or not. Nothing in my mental world seems to be sealed off from my behaviour.

If we think there is causal interaction, then again we seem to be looking for an extension of monist physics rather than a dualism.

Yet it won’t quite do, will it, to say that the physics is all there is to it?

My view is that in fact what’s going on is that we are addressing a question which physics cannot explain, not because physics is faulty or inadequate, but because the question is outside its scope. In terms of physics, we’ve got a type 3 problem; in terms of metaphysics, I hope it’s type 2, though there are some rather discouraging arguments that suggest things are worse than that.

I think the element of mystery in conscious experience is in fact its particularity, its actual reality. All the general features can be explained at a theoretical level by physics, but not why this specific experience is real and being had by me. This is part of a more general mystery of reality, including the questions of why the world is like this in particular and not like something else, or like nothing. We try to naturalise these questions, typically by suggesting that reality is essentially historical, that things are like this because they were previously like that, so that the ultimate explanations lie in the origin of the cosmos, but I don’t think that strategy works very well.

There only seem to be two styles of explanation available here. One is the purely rational kind of reasoning you get in maths. The other is empirical observation. Neither is any good in this context; empirical explanations simply defer the issue backwards by explaining things as they are in terms of things as they once were. There’s no end to that deferral. A priori logical reasoning, on the other hand, delivers only eternal truths, whereas the whole point about reality and my experience is that it isn’t fixed and eternal; it could have been otherwise. People like Stephen Hawking try to deploy both methods, using empirical science to defer the ultimate answer back in time to a misty primordial period, a hypothetical land created by heroic backward extrapolation, where it is somehow meant to turn into a mathematical issue, but even if you could make that work I think it would be unsatisfying as an explanation of the nature of my experience here and now.

I conclude that to deal with this properly we really need a different way of thinking. I fear it might be that all we can do is contemplate the matter and hope pre- or post-theoretical enlightenment dawns, in a sort of Taoist way; but I continue to hope that eventually that one weird trick of metaphysical argument that cracks the issue will occur to someone, because like anyone brought up in the western tradition I really want to get it all back to territory where we can write out the rules and even do some maths!

As I’ve said, this all raises another question, namely why we bother about monism versus dualism at all. Most people realise that there is no single account of the world that covers everything. Besides concrete physical objects we have to consider the abstract entities; those dealt with in maths, for example, and many other fields. Any system of metaphysics which isn’t intolerably flat and limited is going to have some features that would entitle us to call it at least loosely dualist. On the other hand, everything is part of the cosmos, broadly understood, and everything is in some way related to the other contents of those cosmos. So we can equally say that any sufficiently comprehensive system can, at least loosely, be described as monist too; in the end there is only one world. Any reasonable theory will be a bit dualist and a bit monist in some respects.

That being so, the pure metaphysical question of monism versus dualism begins to look rather academic, more about nomenclature than substance. The real interest is in whether your dualism or your monism is any good as an elegant and effective explanation. In that competition materialism, which we tend to call monist, just looks to be an awfully long way ahead.

The Map of Feelings

Posted on by
4

An intriguing study by Nummenmaa et al (paper here) offers us a new map of human feelings, which it groups into five main areas; positive emotions, negative emotions, cognitive operations, homeostatic functions, and sensations of illness. The hundred feelings used to map the territory are all associated with physical regions of the human body.

The map itself is mostly rather interesting and the five groups seem to make broad sense, though a superficial look also reveals a few apparent oddities. ‘Wanting’ here is close to ‘orgasm’. For some years now I’ve wanted to clarify the nature of consciousness; writing this blog has been fun, but dear reader, never  quite like that. I suppose ‘wanting’ is being read as mainly a matter of biological appetites, but the desire and its fulfilment still seem pretty distinct to me, even on that reading.

Generally, a list of methodological worries come to mind, many of which are connected with the notorious difficulties of introspective research. ‘Feelings’ is a rather vaguely inclusive word, to begin with. There are a number of different approaches to classifying the emotions already available, but I have not previously encountered an attempt to go wider for a comprehensive coverage of every kind of feeling. It seems natural to worry that ‘feelings’ in this broad sense might in fact be a heterogeneous grouping, more like several distinct areas bolted together by an accident of language; it certainly feels strange to see thinking and urination, say, presented as members of the same extended family. But why not?

The research seems to rest mainly on responses from a group of more than 1000 subjects, though the paper also mentions drawing on the NeuroSynth meta-analysis database in order to look at neural similarity. The study imported some assumptions by using a list of 100 feelings, and by using four hypothesized basic dimensions – mental experience, bodily sensation, emotion, and controllability. It’s possible that some of the final structure of the map reflects these assumptions to a degree. But it’s legitimate to put forward hypotheses, and that perhaps need not worry us too much so long as the results seem consistent and illuminating. I’m a little less comfortable with the notion here of ‘similarity’; subjects were asked to put feelings closer the more similar they felt them to be, in two dimensions. I suspect that similarity could be read in various ways, and the results might be very vulnerable to priming and contextual effects.

Probably the least palatable aspect, though, is the part of the study relating feelings to body regions. Respondents were asked to say where they felt each of the feelings, with ‘nowhere’, ‘out there’ or ‘in Platonic space’ not being admissible responses. No surprises about where urination was felt, nor, I suppose, about the fact that the cognitive stuff was considered to be all in the head. But the idea that thinking is simply a brain function is philosophically controversial, under attack from, among others, those who say ‘meaning ain’t in the head’, those who champion the extended brain (if you’re counting on your fingers, are you still thinking with just your brain?), those who warn us against the ‘mereological fallacy’, and those like our old friend Riccardo Manzotti, who keeps trying to get us to understand that consciousness is external.

Of course it depends what kind of claim these results might be intended to ground. As a study of ‘folk’ psychology, they would be unobjectionable, but we are bound to suspect that they might be called in support of a reductive theory. The reductive idea that feelings are ultimately nothing more than bodily sensations is a respectable one with a pedigree going back to William James and beyond; but in the context of claims like that a study that simply asks subjects to mark up on a diagram of the body where feelings happen is begging some questions.