Thoughts on Emergence

David J. Chalmers

From: [email protected] (David Chalmers)
Subject: Thoughts on emergence
Date: 6 Oct 90 07:08:47 GMT

Here are some thoughts on "emergence". Nothing definitive, but an attempt to get at the psychological core (or cores) of the notion. Thanks are due to others for providing a stimulating discussion.

Emergence is a tricky concept. It's easy to slide it down a slippery slope, and turn it into something implausible and easily dismissable. But it's not easy to delineate the interesting middle ground in between. Two unsatisfactory definitions of emergence, at either end of the spectrum:

(1) Emergence as "inexplicable" and "magical". This would cover high-level properties of a system that are simply not deducible from its low-level properties, no matter how sophisticated the deduction. This view leads easily into mysticism, and there is not the slightest evidence for it (except, perhaps, in the difficult case of consciousness, but let's leave that aside for now). All material properties seem to follow from low-level physical properties. Very few sophisticated people since the 19th century have actually believed in this kind of "emergence", and it's rarely what is referred to by those who invoke the term favourably. But if you mention "emergence", someone inevitably interprets you as meaning this, causing no end of confusion.

(2) Emergence as the existence of properties of a system that are not possessed by any of its parts. This, of course, is so ubiquitous a phenomenon that it's not deeply interesting. Under this definition, file cabinets and decks of cards (not to mention XOR gates) have plenty of emergent properties - so this is surely not what we mean.

The challenge, then, is to delineate a concept of emergence that falls between the deeply implausible (1) and the overly general (2). After all, serious people do like do use the term, and they think they mean something interesting by it. It probably will help to focus on a few core examples of "emergence":

Note that in all these cases, the "emergent" properties are in fact deducible (perhaps with great difficulty) from the low-level properties (perhaps in conjunction with knowledge of initial conditions), so a more sophisticated concept than (1) is required. Another stab at a definition might be:

(3) Emergent = "deducible but not reducible". Biological and psychological laws and properties are frequently said not to be reducible to physical laws and properties. For many reasons, not the least being that the high-level laws/properties in question might be found associated with all kinds of different physical laws/properties as substrates. (A universe without protons and electrons might nevertheless include learning and memory.)

There are some problems with this definition, though. Firstly, it's not clear what is gained by trying to explicate emergence in terms of the almost-equally-murky concept of "reduction". Secondly, it seems to let in some not-paradigmatically-emergent phenomena, and it's not clear how some emergent phenomena like (A) or (C) would fit this definition. I think that (3) picks out a very interesting class, but it's not quite the class we're after. It's on the right track, though, I think.

The notion of reduction is intimately tied to the ease of understanding one level in terms of another. Emergent properties are usually properties that are more easily understood in their own right than in terms of properties at a lower level. This suggests an important observation: Emergence is a psychological property. It is not a metaphysical absolute. Properties are classed as "emergent" based at least in part on (1) the interestingness to a given observer of the high-level property at hand; and (2) the difficulty of an observer's deducing the high-level property from low-level properties. The properties of XOR are an obvious consequence of the properties of its parts. Emergent properties aren't. Might as well give this a number:

(4) Emergent high-level properties are interesting, non-obvious consequences of low-level properties.

This still can't be the full story, though. Every high-level physical property is a consequence of low-level properties, usually non-obviously. It feels unsatisfactory, for instance, to say that computations performed by a COBOL program are an emergent property relative to the low-level circuit operations - at least this feels much less "emergent" than a connectionist network. So something is missing. The trouble seems to lie with the complex, kludgy organization of the COBOL circuits. The low-level stuff may be simple enough, but all the complexity of the high-level behaviour is due to the complex structure that is given to the low-level mechanisms (by programming). Whereas in the case of connectionism or the game of life it feels that we have simplicity in both low-level mechanisms and their organization. So in those cases, we have much more of a "something for nothing" feel. Let's try for another number:

(5) Emergence is the phenomenon wherein complex, interesting high-level function is produced as a result of combining simple low-level mechanisms in simple ways.

I think this is much closer to a good definition of emergence. Note that COBOL programs, and many biological systems, are excluded by the requirement that not only the mechanisms but their principles of combination be simple. (Of course simplicity, complexity and interestingness are psychological concepts, at least for now, though we might try to explicate them in terms of Chaitin-Kolmogorov-Solomonoff complexity if we felt like it. My intuition is that this is likely to prove a little simplistic, although Chaitin has an interesting paper that attempts to derive a notion of the "organization" of a system using similar considerations.) And note also that most things that satisfy this definition should also satisfy (4) - due to our feeling that simple principles should have simple consequences (or else complex but uninteresting consequences, like random noise). Any complex, interesting consequence is likely to be non-obvious.

This does indeed fit in with the feeling that emergence is a "something for nothing" phenomenon - though in a more subtle and satisfactory way than set forth in (1), for instance. It's a phenomenon whereby "something stupid buys you something smart". And most of our examples fit. The game of Life and connectionist networks are obvious: interesting high-level behaviour as a consequence of simple dynamic rules for low-level cell dynamics. In evolution, the genetic mechanisms are very simple, but the results are very complex. (Note that there is a small difference, in that in the latter case the emergence is diachronic, i.e. over time, whereas in the first two cases the emergence is synchronic, i.e. not over time but over levels present at a given time.)

We're still not completely there - it's not clear how (C), the operating system example, fits into this paradigm of emergence. But throwing in a smidgen of teleology should get us the rest of the way. I.e., we have to notice that everything here has to be relativized to design. So we design the game of Life according to certain simple principles, but complex, interesting properties leap out and surprise us. Similarly for the connectionist network - we only design it at a low level (though in this case we hope that complex high-level properties will emerge). Whereas in the COBOL case - and in the case of much traditional AI - you only get out what you put in (N.B. I'm not necessarily knocking this: at least here, I'm trying to explicate emergence, not to defend it). And now the operating system example fits in well. The design principles of the system in this case are quite complex - unlike the other cases that fit (5) above - but still the figure "35" is not a part of that design at all. So:

(6) Emergence is the phenomenon wherein a system is designed according to certain principles, but interesting properties arise that are not included in the goals of the designer.

Notice the appearance of the word "goal" - this is important, any design is goal-relative. So the notion now is quite teleological. I notice that Russ Abbott makes a similar point in a recent posting. Notice, however, that as we've conceded that emergence is a psychological property, we're able to construe teleology in a psychological, non-absolute way. So for our purposes here, we only need the appearance of teleology. This is nice, because it allows us to include system where strictly speaking, "design" doesn't apply at all. In evolution, for instance, there is no "designer", but it is easy to treat evolutionary processes as processes of design. On more than one level.

We can view evolution as teleological at the level of the gene - as in Dawkins' theory, for instance. Then the appearance of complex, interesting high-level properties such as intelligence is quite emergent. We also can reconstrue evolution as teleological at the level of the organism (this is perhaps a more straightforward Darwinian view of things). On this construal, the most salient adaptive phenomena like intelligence are no longer emergent, but the goal of the design process. However, this view does open up the possibility of other kinds of emergent phenomena: firstly, non-selected-for byproducts of the evolutionary process (such as Gould and Lewontin's "Spandrels"); secondly and more intriguingly, it allows an explanation for why "consciousness" (or "subjectivity" or "qualia" or whatever) seems emergent. Raw consciousness doesn't not seem to have been selected for, as it doesn't play any direct functional role (though it does have functional counterparts; this is a subtle issue, but remember we're talking about the way things seem, not the way they are); but it somehow emerges as a byproduct of selection for adaptive process such as intelligence.

It's probably foolish to search for a definitive construal of "emergence": like most psychological concepts, it probably is best construed as a "family resemblance" - each of the "definitions" outlined above might play some role. Personally, I'm happiest with a combination of (5) and (6) - with (5) being the "core" variety of emergence, and (6) being a more general variety of which (5) is a special case.

From: [email protected] (David Chalmers)
Subject: Re: Thoughts on emergence
Date: 8 Oct 90 22:56:22 GMT

In article <[email protected]> [email protected] (cameron shelley) writes:

It depends on how you construe the teleology of the evolutionary process. Of course there's no real teleology there, but psychologically we can construe it an at least three different ways. My original post suggested two of these ways, your response suggests a third.

First way: Teleology at the level of the gene. Goal: get genes to replicate. At this level, intelligence is emergent.

Second way: Teleology at the level of the organism. Goal: get organisms to survive and replicate. This is your suggestion. If the only teleology we impute to the evolutionary process is that of a "natural selector" - i.e. produce systems that survive and replicate - then specific functions like intelligence are indeed emergent.

Third way: Teleology at the level of the organism. Goal: get organisms to be like X (for some X, e.g. X = strong, fast, intelligent...). Here, we are construing the teleology of the evolutionary process not as that of the natural selector, but as that of the "blind watchmaker". On this view, intelligence is the goal of the evolutionary process, and so cannot be regarded as emergent.

The teleology here not being metaphysical but psychological, all these ways of construing it are quite valid. My original post only mentioned the first and third possibilities. The second possibility is also a very reasonable construal, and serves the useful purpose of showing how intelligence can be regarded as emergent without having to descend to the level of the "selfish gene". Sorry for omitting this possibility in my original post.