Sergio’s Computational Functionalism

sergio differenceSergio has been ruminating since the lively discussion earlier and here by way of a bonus post, are his considered conclusions…..

Linespace

 

Not so long ago I’ve enthusiastically participated in the first phases of the discussion below Peter’s post on “Pointing”. The discussion rapidly descended in the fearsome depths of the significance of computation. In the process, more than one commentator directly and effectively challenged my computationalist stance. This post is my attempt to clarify my position, written with a sense of gratitude to all: thanks to all for challenging my assumptions so effectively, and to Peter for sparking the discussion and hosting my reply.

This lengthy post will proceed as follows: first, I’ll try to summarise the challenge that is being forcefully proposed. At the same time, I’ll explain why I think it has to be answered. The second stage will be my attempt to reformulate the problem, taking as a template a very practical case that might be uncontroversial. With the necessary scaffolding in place, I hope that building my answer will become almost a formality. However, the subject is hard, so wish me luck because I’ll need plenty.

The challenge: in the discussion, Jochen and Charles Wolverton showed that “computations” are arbitrary interpretations of physical phenomena. Because Turing machines are pure abstractions, it is always possible to arbitrarily define a mapping between the evolving states of any physical object and abstract computations. Therefore asking, “what does this system compute?” does not admit a single answer, the answer can be anything and nothing. In terms of one of our main explananda: “how do brains generate meanings?” the claim is that answering “by performing some computation” is therefore always going to be an incomplete answer. The reason is that computations are abstract: physical processes acquire computational meaning only when we (intentional beings) arbitrarily interpret these processes in terms of computation. From this point of view, it becomes impossible to say that computations within the brain generate the meanings that our minds deal with, because this view requires meanings to be a matter of interpretation. Once one accepts this point of view, meanings always pre-exist as an interpretation map held by an observer. Therefore, “just” computations, can only trade pre-existing (and externally defined!) meanings and it would seem that generating new meanings from scratch entails an infinite regression.

To me, this is nothing but another transformation of the hard problem, the philosophical kernel that one needs to penetrate in order to understand how to interpret the mechanisms that we can study scientifically. It is also one of the most beautifully recursive problem that I can envisage: the challenge is to generate an interpretative map that can be used to show how interpretative maps can be generated from scratch, but this seems impossible, because apparently you can only generate a new map if you can ground it on a pre-existing map. Thus, the question becomes: how do you generate the first map, the first seed of meaning, a fixed reference point, which gets the recursive process started?

In the process of spelling out his criticism, Jochen gave the famous example of a stone. Because the internal state of the stone changes all the time, for any given computation, we can create an ad-hoc map that specifies the correspondence of a series of computational steps with the sequence of internal steps of our stone. Thus, we can show that the stone computes whatever we want, and therefore, if we had a computational reduction of a mind/brain, we could say that the same mind exists within every stone. Consequently, computationalism can either require some very odd form of panpsychism or be utterly useless: it can’t help to discriminate between what can generate a mind and what can’t. I am not going to embrace panpsychism, so I am left with the only option of biting the bullet and show how this kind of criticism can be addressed.

Without digressing too much, I hope that the above leaves no doubt about where I stand: first, I think this critique of computational explanations of the (expected) mind/brain equivalence is serious, it needs an answer. Furthermore, I also think that answering it convincingly would count as significant progress, even a breakthrough, if we take ‘convincingly’ to stand for ‘capable of generating consensus’. Dissolving this apparently unsolvable conundrum is equivalent to showing why a mechanism can generate a mind, I don’t know if there is a bigger prize in this game.

I’ll start from my day job, I write software for a living. What I do is write instructions that would make a computer reliably execute a given sequence of computations, and produce the desired results. It follows that I can, somehow, know for sure what computations are going to be performed: if I couldn’t, writing my little programs would be vain. Thus, there must be something different between our ordinary computers and any given stone. The obvious difference is that computers are engineered, they have a very organised structure and behaviour, specifically because this makes programming them feasible. However, in theory, it would be possible to produce massively complicated input/output systems to substitute the relevant parts (CPU, RAM, long-term memory) of a computer with a stone, we don’t do this because it is practically far too complicated, not because it is theoretically impossible. Thus, the difference isn’t in the regular structure and easily predictable behaviour of the Von Neumann/Harvard and derived architectures. I think that the most notable differences are two:

  1. When we use a computer, we already have agreed upon the correct way to interpret its output. More specifically, all the programs that are written assume such a mapping, and would produce outputs that conform to it. If a given program will be used by humans (this isn’t always the case!) the programmer will make sure that the results will be intelligible to us. Similarly, the mapping between the computer states and their computational meaning is also fixed (so fixed and agreed in advance, that I don’t even need to know how it works, in practice).
  2. In turn, because the mapping isn’t arbitrary, also the input/output transformations follow predefined and discrete sets of rules. Thus, you can plug different monitors and keyboards, and expect them to work in similar ways.

For both differences, it’s a matter of having a fixed map (we can for simplicity collapse the maps from 1 & 2 into a single one). Once our map is defined and agreed upon, we can solve the stone problem and say “computer X is running software A, computer Y is running software B” and expect everyone to agree. The arbitrariness of the map becomes irrelevant because in this case the map itself has been designed/engineered and agreed from the start.

This isn’t trivial, because it becomes enlightening when we propose the hypothesis that brains can be modelled as computers. Note my wording: I am not saying “brains are computers”, I talk about “modelled” because the aim is to understand how brains work, it’s an epistemological quest. We are not asking “what brains/minds are”; in fact, I’ll do all I can to steer away from ontology altogether.

Right, if we assume that brains can be modelled as computers, it follows that it should be possible to compose a single map that would allow us to interpret brain mechanisms in terms of computations. Paired with a perfect brain scanner (a contraption that can report all of the brain states that are required to do the mapping) such a map would allow us to say without doubt “this brain is computing this and that”. As a result, with relatively little additional effort, it should become possible to read the corresponding brain. From this point of view, the fact that there is an infinite number of possible maps, but only one is “the right” one, means that the problem is not about arbitrariness (as it seemed for the stone). The problem is entirely different, it is about finding the correct map, the one that is able to reliably discern what the scanned mind is thinking about. This is why in the original discussion I’ve said the arbitrariness of the mapping is the best argument for a computational theory of the mind. It ensures the search space for the map is big enough to give us hope that such a map does exist. Note also that all of the above is nothing new, it is just stating explicitly the assumptions that underline all of neuroscience; if there are some exceptions, they would be considered very unorthodox.

However, this where I think that the subject becomes interesting. All of the above has left out the hard side of the quest, I haven’t even tried to address the problem of how computations can generate a “meaningful map” on its own. To tackle this mini-hard problem, we need to go back to where we started and recollect how I’ve described the core of the “anti-computalism” stance. Taking about brain/mechanisms, I’ve asked: how [does the brain] generate the first map, the first seed of meaning, a fixed reference point, which gets the recursive process started? Along the way, I’ve claimed that it is reasonable to expect that a different but important map can be found, the one that describes (among many other things) how to translate brain events into mind events (thoughts, memories, desires, etc.). Therefore, one has to admit that this second map (our computational interpretation) would have to contain, at least implicitly, the answer on the fixed reference point. How is this possible? Note that I’ve strategically posed the question in my own terms, and mentioned the need for a fixed reference point. You may want to recall the “I-token” construct of Retinoid Theory, but in general, one can easily point out that the reference point is provided by the physical system itself. We have, ex-hypothesis, a system that collects “measurements” from the environment (sensory stimuli), processes them, and produces output (behaviour); this output is usually appropriate to preserve the system integrity (and reproduce, but that’s another story). Fine, such a system IS a fixed reference point. The integrity that justifies the whole existence of the system IS precisely what is fixed – all the stimuli it collects are relative to the system itself. As long as the system is intact enough to function, it can count as a fixed reference point; with a fixed reference, meanings become possible because reliable relations can be identified, and if they can, then they can be grouped together and produce more comprehensive “interpretative” maps. This is the main reason why I like Peter’s Haecceity: it’s the “thisness” of a particular computational system that actually seeds the answer of the hard side of the question.

Note also that all of the above captures the differences I’ve spelled out between a standard computer and a common stone. It’s the specific physicality of the computer that ultimately distinguishes it from a stone: in this case, we (humans) have defined a map (designing it from scratch with manageability in mind) and then used the map to produce a physical structure that will behave accordingly. In the case of brains/minds, we need to proceed in the opposite direction: given a structure and its dynamic properties, we want to define a map that is indeed intelligible.

Conclusions:

  • The computational metaphor should be able to capture the mechanisms of the brain and thus describe the (supposed) equivalence between brain events and mind events.
  • Such description would count as a weak explanation as it spells out a list of “what” but doesn’t even try to produce a conclusive “why”.
  • However, just expecting such mapping to be possible already suggests where to find the “why” (or provides it, if you feel charitable). If such mapping will prove to be possible, it follows that to be conscious, an entity needs to be physical. Its physicality is the source of the ability of generating its own, subjective meanings.
  • This in turns reaffirms why our initial problem, posed by the unbounded arbitrariness of computational explanations, does not apply. The computational metaphor is a way to describe (catalogue) a bunch of physical processes, it spells out the “what” but is mute on the “why”. The theoretical nature of computation is the reason why it is useful, but also points to the missing element: the physical side.
  • If such a map will turn out to be impossible, the most likely explanation is that there is no equivalence between brain and mind events.

 

Finally, you may claim that all these conclusions are themselves weak. Even if the problematic step of introducing Haecceity/physicality, as the requirement to bootstrap meaning, is accepted, the explanation we gain is still partial. This is true, but entails the mystery of reality (again, following Peter): because cognition can only generate and use interpretative maps (or translation rules), it “just” shuffles symbols around, it cannot, in no way or form ultimately explain why the physical world exists (or what exactly the physical world is, this is why I steered away from ontology!). Because all knowledge is symbolic, some aspect of reality always has to remain unaccounted and unexplained. Therefore, all of the above can still legitimately feel unsatisfactory: it does not explain existence. But hey, it does talk about subjectivity and meaning (and by extension, intentionality), so it does count as (hypothetical) progress to me.

Now please disagree and make me think some more!

Dancing Pixies

Picture: Dancing Pixies. I see that among the first papers published by the  recently-launched journal of Cognitive Computation, they sportingly included one arguing that we shouldn’t be seeing cognition as computational at all.  The paper, by John Mark Bishop of Goldsmith’s, reviews some of the anti-computational arguments and suggests we should think of cognitive processes in terms of communication and interaction instead.

The first two objections to computation are in essence those of Penrose and Searle, and both have been pretty thoroughly chewed over in previous discussions in many places; the first suggests that human cognition does not suffer the Gödelian limitations under which formal systems must labour, and so the brain cannot be operating under a formal system like a computer program; the second is the famous Chinese Room thought experiment. Neither objection is universally accepted, to put it mildly,  and I’m not sure that Bishop is saying he accepts them unreservedly himself – he seems to feel that having these popular counter-arguments in play is enough of a hit against computationalism in itself to make us want to look elsewhere.

The third case against computationalism is the pixies: I believe this is an argument of Bishop’s own, dating back a few years, though he scrupulously credits some of the essential ideas to Putnam and others.  A remarkable feature of the argument is that uses panpsychism in a reductio ad absurdum (reductio ad absurdum is where you assume the truth of the thing you’re arguing against, and then show that it leads to an absurd, preferably self-contradictory conclusion).

Very briefly, it goes something like this; if computationalism is true, then anything with the right computational properties has true consciousness (Bishop specifies Ned Block’s p-consciousness, phenomenal consciousness, real something-that-it-is-like experience). But a computation is just a given series of states, and those states can be indicated any way we choose.  It follows that on some interpretation, the required kind of series of states are all over the place all the time. If that were true, consciousness would be ubiquitous, and panpsychism would be true (a state of affairs which Bishop represents as being akin to a world full of pixies dancing everywhere). But since, says Bishop, we know that panpsychism is just ridiculous, that must be wrong, and it follows that our initial assumption was incorrect; computationalism is false after all.

There are of course plenty of people who would not accept this at all, and would instead see the whole argument as just another reason to think that panpsychism might be true after all. Bishop does not spend much time on explaining why he thinks panpsychism is unacceptable, beyond suggesting that it is incompatible with the ‘widespread’ desire to explain everything in physical terms, but he does take on some other objections more explicitly.  These mostly express different kinds of uneasiness about the idea that an arbitrary selection of things could properly constitute a computation with the right properties to generate consciousness.

One of the more difficult is an objection from Hofstadter that the required sequences of states can only be established after the fact: perhaps we could copy down the states of a conscious experience and then reproduce them, but not determine them in advance. Bishop uses an argument based on running the same consciousness program on a robot twice; the first time we didn’t know how it would turn out; the second time we did (because it’s an identical robot and identical program) but it’s absurd to think that one run could be conscious and the other not. 

Perhaps the most tricky objection mentioned is from Chalmers; it points out that cognitive processes are not pre-ordained linear sequences of states, but at every stage have the possibility of branching off and developing differently. We could, of course remove every conditional switch in a given sequence of conscious cognition and replace it by a non-conditional one leading on to the state which was in fact the next one chosen. For that given sequence, the outputs are the same – but we’re not entitled to presume that consious experience would arise in the same way because the functional organisation is clearly different, and that is the thing, on computationalist reasoning, which needs to be the same.  Bishop therefore imagine a more refined version: two robots run similar programs; one program has been put through a code optimiser which keeps all the conditional branches but removes bits of code which follow, as it were, the unused branches of the conditionals. Now surely everything relevant is the same: are we going to say that consciousness arises in one robot by virtue of there being bits of extra code there which lie there idle? That seems odd.

That argument might work, but we must remember that Bishop’s reductio requires the basics of consciousness to be lying around all over the place, instantiated by chance in all sorts of things. While we were dealing with mere sequences of states, that might look plausible, but if we have to have conditional branches connecting the states (even ones whose unused ends have been pruned) it no longer seems plausible to me.  So in patching up his case to respond to the objection, Bishop seems to me to have pulled out some of the foundations it was originally standing on. In fact, I think that consciousness requires the right kind of causal relations between mental states, so that arbitrary sets or lists of states won’t do.

The next part of the discussion is interesting. In many ways computationalism looks like a productive strategy, concedes Bishop – but there are reasons to think it has its limitations. One of the arguments he quotes here is the Searlian point that there is a difference between a simulation and reality. If we simulate a rainstorm on a computer, no-one expects to get wet; so if we simulate the brain, why should we expect consciousness? Now the distinction between a simulation and the real thing is a relevant and useful one, but the comparison of rain and consciousness begs the question too much to serve as an argument. By choosing rain as the item to be simulated, we pick something whose physical composition is (in some sense) essential; if it isn’t made of water it isn’t rain. To assume that the physical substrate is equally essential for consciousness is just to assume what computationalism explicitly denies.  Take a different example; a letter. When I write a letter on my PC, I don’t regard it as a mere simulation, even though no paper need be involved until it’s printed; in fact, I have more than once written letters which were evntually sent as email attachments and never achieved physical form. This is surely because with a letter, the information is more essential than the physical instantiation. Doesn’t it seem highly plausible that the same might be true to an even greater extent of consciousness? If it is true, then the distinction between simulation and reality ceases to apply.

To make sceptical simulation arguments work, we need a separate reason to think that some computation was more like a simulation than the reality – and the strange thing is, I think that’s more or less what the objections from Hofstadter and Chalmers were giving us; they both sort of draw on the intuition that a sequence of states could only simulate consciousness  in the sort of way a series of film frames simulates motion.

The ultimate point, for Bishop, is to suggest we should move on from the ‘metaphor’ of computation to another based on communication. It’s true that the idea of computation as the basis of consciousness has run into problems over recent years, and the optimism of its adherents has been qualified by experience; on the other hand it still has some remarkable strengths. For one thing, we understand computation pretty clearly and thoroughly;  if we could reduce consciousness to computation, the job really would be done; whereas if we reduce consciousness to some notion of communication which still (as Bishop says) requires development and clarification, we may still have most of the explanatory job to do.

The other thing is that computation of some kind, if not the only game in town, still is far closer to offering a complete answer than any other hypothesis.  I supect many people who started out in opposing camps on this issue would agree now that the story of consciousness is more likely to be ‘computation plus plus’ (whatever that implies) than something quite unrelated.