learning

Simon Armitage and colleagues [1] describe archaeological remains from Jebel Faya, in the United Arab Emirates. The assemblages come from a rock shelter in the mountain, which is around 100 km south of the Straits of Hormuz, entry to the Persian Gulf. Below Bronze Age and later remains, are three Paleolithic units. The oldest (assemblage C in the paper) is dated by OSL to the last interglacial, around 125,000 years ago. My comments here are more note-like than usual; this topic opens a window into some work we've been recently doing.

The authors' main conclusion is that the oldest assemblage displays technical similarities to East African archaeological assemblages, which are not present in the archaeology of the Levant either before or after this time. We have to dig into the supplementary material to the paper to get a good account of the technical similarities:

Technologically, this assemblage has general links to East Africa (S3 S4) while showing none of the technological traits characteristic in the contemporaneous Levantine Mousterian (S5). As in the early Middle Stone Age (MSA) of East Africa, Assemblage C exhibits three profoundly different reduction strategies: bifacial, volumetric blade, and radial Levallois. This combination is unknown in the Levant after about 200 ka, where there is no bifacial reduction and the Levallois method is largely limited to unidirectional converging. The latter produced large numbers of Levallois points, which are absent from Assemblage C.

For a layman's description of the result from coauthor Anthony Marks, I can recommend Katherine Harmon's account at Scientific American's website.

I like the observation, but I think we should be cautious about it. The basic idea is that African assemblages display three different strategies early in the reduction sequence, none of which are evident in Levantine assemblages of equivalent age.

Reduction sequences and conservatism

Yesterday I talked over this concept with my graduate student Marc Kissel. I find it very interesting that the authors focused on initial reduction stages as elements of technical similarity. They thereby assume much about the cultural transmission of the reduction sequence.

It seems reasonable that the initial steps of a reduction sequence -- from quarrying through early core shaping -- should be conservative. Early stages necessarily constrain the later steps toward finished tool production, so that a skilled toolmaker who wants to carry out the later stages of a reduction sequence has first to get the early steps right.

Paradoxically a naive learner may be ill-equipped to attend to the importance of these first steps, compared to later steps where the preform is more readily identifiable by its physical configuration. Within a social group, the early steps of reduction may well be carried out by other people, including less-skilled artificers. The best toolmaker may go to the quarry himself, but often he may call on someone less skilled to carry out the initial reduction, or may be forced to work with partially exhausted cores from earlier attempts.

I'm willing to hazard a guess that the social learning that enables tool manufacture would exert a bias toward low error rates early in the reduction sequence. We can consider a biological analogy -- early embryonic development is more strongly conserved across taxa (and phyla) than later development. Changing something early in a developmental sequence may make later events impossible. If I'm right, the argument by Armitage and colleagues should have some force -- finding that the early stages of the reduction sequence are shared among sites should be a better indicator of relationship than most archaeological indicators.

But Armitage and colleagues' conclusion has force just to the extent that we accept two proposals: (1) that we understand the technical variation in the Levant, and (2) that independent development of the early-stage reduction strategies in the Jebel Faya assemblage is unlikely.

These proposals hang together. The Levant is richly documented across the period before and after the last interglacial, moreso after OIS 6 (around 130,000 years ago) than before. These assemblages were directed toward convergent removal of Levallois points. I'm not immediately in a position to discuss the variation within these assemblages, but the question strikes me as crucial. Although the archaeological record from this area is relatively dense, like all places it samples only a small fraction of the actual groups that must have existed at the time -- to use a genetic comparison, the record has high coverage over a very small fraction of the regional behaviorome.

Was independent invention of these early-stage reduction strategies likely? The answer depends on whether a particular early-stage reduction strategy is merely rare in the large Levantine sample, or entirely absent. If such a strategy (in this case, foliate reduction) occurs at all, we can infer that its invention was possible, if not likely. With assemblage C at Jebel Faya, we are considering the cultural tradition represented by 500 artifacts. If we treated these as a random sample of the Levantine record, they are exceedingly unusual, no doubt. But random sampling across an entire record isn't the correct comparison; we want some equivalent sampling of the cultural information in terms of time and space.

The paper's conclusion that Jebel Faya represents an incursion of African-derived technical traditions into the Arabian peninsula depends on these assertions. I don't have strong feelings about them, but I think we should work to get a better statistical understanding about the issue. I am singularly unimpressed when archaeologists assert that one assemblage "resembles" another on purely typological grounds. Typological similarities may result from many constraints other than cultural information, and rare appearances actually carry a lot of information about them.

Out of Africa early

Now, what about this "southern route" business? I say it's a year behind the times. The entire reason for the "southern route" hypothesis was to explain how Africans could have left Africa 70,000 years ago without being stopped by Neandertals in the Levant. Sail them around the southern coast of Asia, and you can get them early into SE Asia and Australia without mixing with those darned Neandertals.

We obviously don't need to rule out Neandertal interbreeding anymore. We know it happened, most likely in West Asia. Putting Africans into the Levant during the last interglacial isn't a bug, it's a feature. We need contact between moderns and Neandertals in this area to explain the genetic data.

The dates may seem like more of a stumbling block. If we accept that a major out-of-Africa movement was underway by 70,000 years ago, we are going to have a hard time explaining why the Levant seems to have been entirely uninfluenced by it.

But a 70,000-year-long chronology, based on estimates of mtDNA haplogroup divergences, is already out of kilter with the majority of evidence. Nuclear DNA suggests a substantially longer timescale, which would derive non-African and sub-Saharan populations from common ancestors before 140,000 years ago. Depending on the amount of mixture among these populations and the mutation rate we adopt, these populations may have begun to differentiate very early in the Middle Stone Age.

It's hard to account for the diversity of people outside of Africa with a short migration timescale. People outside Africa are around 20 percent more inbred than sub-Saharan Africans, but they don't look like they underwent any sudden severe bottleneck. Even accounting for the mixture with archaic people like Neandertals and Denisovans, much of the variation of Middle Pleistocene humans (still present in Africa) just didn't get into non-Africans.

I would propose a movement of MSA Africans into West Asia before the last interglacial as a model that provides a good fit to these data. An early movement followed by long interactions in this limited area would explain so much of the population structure and morphological variation of MSA Africans wasn't represented in the people who peopled Eurasia. A substantial delay between the initial entrance into West Asia and the dispersal to Europe and the rest of Asia would explain why the later archaeological transitions in those regions have no sign of immediate technical or cultural links to the MSA. It would also explain why the initial "modern" humans outside Africa share few if any derived morphological features with Africans after 100,000 years ago.

The anatomy of the Skhul and Qafzeh samples suggests that an African incursion into the Near East did occur before 100,000 years ago. Many paleoanthropologists have supposed that this early incursion did not persist, even locally. The later Levantine sample includes individuals with more Neandertal resemblances, chiefly Amud and Kebara. But each of the later specimens shares several traits with early modern humans from Skhul or Qafzeh. Indeed there is no clear constellation of derived traits that sorts the Skhul-Qafzeh sample cleanly from Tabun 1 and the later Levantine specimens. I just don't think this skeletal record poses any problem for the idea of a long interaction of populations in this area -- especially if we extend the focus from the Levant into the Arabian peninsula and Persian Gulf region.

The strongest reason to suppose that an African incursion was extinguished is not the skeletal record but instead the mtDNA timescale. I can refer readers to the paper by Endicott and colleagues [2], which discusses a range of mutation rate estimates and their effects on the origin of macrohaplogroups M and N, the key ancestral non-African lineages. Current estimates unanimously suggest that these clades originated within the last 75,000 years. By itself, this would suggest that the mtDNA common ancestors of non-Africans and sub-Saharan African populations diverged shortly before that time.

I keep coming back to this, because the mtDNA just seems so out of line with the autosomal and X-chromosome picture. I regard this as a serious sticking point and hesitate to just wave it away. As I suggested to Charles Choi, the resolution may involve a time of isolation outside Africa during which the ancestors of non-Africans lost heterozygosity (and became enriched for the later mtDNA clades M and N). Or maybe we just have the mtDNA clock wrong -- the large revisions of the Neandertal-human mtDNA divergence in the light of developing evidence don't inspire confidence about the timing of internal nodes to the human mtDNA tree.

The early archaeological assemblage from Jebel Faya strikes me as consistent with a model of early dispersal from Africa, but not especially good evidence for it. The outstanding question is whether the early reduction strategy is a behavioral trait that provides good evidence about biological relationships. I see the logic but think that it is tenuous.

The model obviously is relevant to the question of an early presence of African-derived modern humans in India. If we combine the presence of an African-derived population in eastern Arabia with the large exposed Persian Gulf region during the last interglacial, this begins to look like a large habitable region with easy land connections to the Indus River valley. But the Indian subcontinent would potentially have been home to a very large population of ancient humans. I doubt that an occupation across the large area of West Asia plus the Indian subcontinent would have enabled the substantial reduction of heterozygosity that we see in present-day non-Africans.

References

Dario Floreano and Laurent Keller describe experiments that combine genetic algorithms and robots. It's a review essay rather than a description of new research, but unlike most descriptions of "evolutionary robotics", it's actually directed toward biologists instead of AI researchers.

In this essay we will examine key experiments that illustrate how, for example, robots whose genes are translated into simple neural networks can evolve the ability to navigate, escape predators, coadapt brains and body morphologies, and cooperate. We present mostly—but not only—experimental results performed in our laboratory, which satisfy the following criteria. First, the experiments were at least partly carried out with real robots, allowing us to present a video showing the behaviours of the evolved robots. Second, the robot's neural networks had a simple architecture with no synaptic plasticity, no ontogenetic development, and no detailed modelling of ion channels and spike transmission. Third, the genomes were directly mapped into the neural network (i.e., no gene-to-gene interaction, time-dependent dynamics, or ontogenetic plasticity). By limiting our analysis to these studies we are able to highlight the strength of the process of Darwinian selection in comparable simple systems exposed to different environmental conditions.

Some of the simplest machine learning experiments are basically like those used in behavioral psychology -- put the robots in a maze, make them remember where the food is, that sort of thing. Robots are simpler than rats, so the researchers can reverse-engineer the "evolved" software at the end of a series of experiments to see what worked and why:

Interestingly, the driving speed of the best-evolved robots was approximately half of the maximum possible speed and did not increase even when the evolutionary experiments were continued for another 100 generations. Additional experiments where the speed was artificially increased revealed that fast-moving robots had high rates of collisions because the 300-ms refresh rate of the sensors did not allow them to detect walls sufficiently in advance at high speed. Thus, the robots evolved to move at intermediate speeds because of their limited neural and sensory abilities.

Figuring out that particular optimization would drive a team of human programmers crazy. Can you imagine? "Why do they keep running into that wall?!"

On the other hand, dumb selection took a lot of generations to get to that point. You can't say selection was more efficient. If you had a crew of programming grunts and forced them to sit in a room for 100 robot generations, they'd come up with something.

It's quite possible that a human would have come up with much better software, by pushing the robots past the limits of mutations on their "genomes". Selection has its own "sensor limitations", it can get stuck in a local optimum, and depends on the mutation structure to explore the landscape.

It helps if the landscape has some strong correlation structure. That's what came to my mind as I read their account of experiments to make robots cooperate:

However, when the arena contained both large and small tokens, the behaviour of robots was influenced by the group kin structure. In groups of unrelated robots (i.e., robots whose genomes where not more similar within than between groups), robots invariably specialised in pushing the small objects, which was the most efficient strategy to maximise their own individual fitness them (i.e., large tokens provided an equal direct payoff as a small token but were more difficult to successfully push). By contrast, the presence of related robots within groups allowed the evolution of altruism. When groups were formed of “clonal” robots all having the same genome, individuals primarily pushed the large tokens even though it was costly, in terms of individual fitness, for the robots pushing (Video S6).

If you wonder how robots have "kin", it's that they share similar (or the same) genomes. The simplicity of the behaviors suggests a functional explanation for kin selection -- for many kinds of tasks, it may simply be easier to cooperate with other individuals who "work" the same way. Different approaches to the same task may clash.

They describe a similar result for cooperation by information sharing:

Similar results were obtained in experiments where groups of light-emitting, foraging robots could communicate the position of a food source at a cost to themselves because of the resulting increased competition near food. In these experiments, robots again readily evolved costly communication when they were genetically related, but altruistic communication never evolved in groups of unrelated robots when selection operated at the individual level [38],[39].

The next logical step for this kind of research is nano-scale: evolutionary robotics on molecular machines. Which is scary. I hope they have the sense not to train them up by eating biological systems...

There's this old course on the books here, "Human aspects of robotics". I suppose it was taught back in the 80's when robots looked like they would replace all the manufacturing workers. I've often thought that someday it may be revived as with robots as the heroes instead of the villains.

References:

Floreano D, Keller L. 2010. Evolution of adaptive behavior in robots by means of Darwinian selection. PLoS Biol 8:e1000292. doi:10.1371/journal.pbio.1000292

I'm a big booster of the idea that human demographic expansion helped drive our recent evolution. So you might expect me to like the new paper by Adam Powell, Stephen Shennan and Mark Thomas, titled, "Late Pleistocene demography and the appearance of modern human behavior." Yet, I see a lot of weaknesses in the paper. I think the paper tries to sidestep several issues about "modern human behavior" that ought to be tackled head-on. In the end, the model in the paper can't describe the data the authors want to consider. Maybe they should have adopted a different model; maybe different data.

I've taken a lot of notes about this -- too many for me to share, but I wanted to review the basic exposition of the paper, including why the authors think demography may determine technological change during the Late Pleistocene. I might post other notes later on the issue of genetic modeling of demography and its relevance for archaeology.

The authors describe a model in which the density of a metapopulation determines the rate of increase (or decline) its cultural evolution, using simulations to extend analytical results from Henrich (2004). Follow their assumptions and you arrive at the conclusion that population density can, under certain conditions, constrain the trajectory of cultural change.

The question is whether the model's assumptions can apply to the real world. Here's the abstract of the paper:

The origins of modern human behavior are marked by increased symbolic and technological complexity in the archaeological record. In western Eurasia this transition, the Upper Paleolithic, occurred about 45,000 years ago, but many of its features appear transiently in southern Africa about 45,000 years earlier. We show that demography is a major determinant in the maintenance of cultural complexity and that variation in regional subpopulation density and/or migratory activity results in spatial structuring of cultural skill accumulation. Genetic estimates of regional population size over time show that densities in early Upper Paleolithic Europe were similar to those in sub-Saharan Africa when modern behavior first appeared. Demographic factors can thus explain geographic variation in the timing of the first appearance of modern behavior without invoking increased cognitive capacity.

You can always tell what's supposed to be bad, it's the thing that you're not supposed to to "invoke". You know, like witches and vampires.

The model

In fact, "cognitive capacity", as a continuous, one-dimensional variable, underlies the model. In a nutshell, the model assumes that people learn behaviors by instantaneously absorbing the "skill" (which I'll call "mojo") from the best (highest "mojo") individual in their population. But they don't learn perfectly; their mojo ends up varying.

Nevertheless the whole population is choosing one individual to copy, so what happens over time is that the population changes in one direction or the other. If the distribution among individuals includes a few with higher mojo, then the average amount of mojo should increase over time. Imagine if the whole population copied the running style of the best 100 m runner. The world record might reduce over time; and then people copy the new world record holder, and the average speeds up again, ad infinitum. There is stochastic variation from one step to the next -- sometimes it will increase more, sometimes less, and sometimes it may shrink a little. But the model is deterministic: depending on the distribution of mojo, it will either trend upward or downward.

I picked the analogy because it points out a weakness of the model. There's no possibility of reaching an optimum, or a stasis. In fact, the survival value of "mojo" simply isn't part of the model, nor is the cost of developing mojo.

OK, it's a simple model -- too simple to capture most aspects of reality. What value can it possibly have?

The assumption is that some behaviors take more mojo than others. Some behaviors then will lie near a threshold where the population is just at the border between gaining or losing mojo over time. The fastest runner in the population might still be slower than last year's champion. If the population models the new winner, they might lose mojo on average.

So the change in mojo doesn't depend on the current average; it depends on the distribution of the highest-mojo individual. That's an extreme value, and extreme values depend on the total number of individuals. There's some chance that the Jamaican national champion will be the Olympic gold medalist -- like last year. But on average the world champion is faster than the champion of any single country; the champion of a country is faster than the champion of any average local track club, and so on. Numbers make a difference. Add more individuals, and you have a better chance of a high extreme value -- a better chance in the model that mojo will increase.

Again, the analogy shows the model's deficiencies. Local track clubs don't vary randomly. There are some local track clubs where the average 100 m time is pretty close to the Olympic champion's. In part this is because information isn't shared instantly and universally. There are both explicit dynamics and path-dependence: Jamaica's running team has been so successful in part because of recent investments in infrastructure, in part because of leadership from a few gifted coaches. And in large part it's because talent matters. Some people just have more running mojo.

But the model does show that for a limited range of behaviors, population size (in Powell and colleagues' simulations, local population density) can exert a deterministic effect on the behavior of the population. Outside that range, the behavior will be dominated by non-demographic factors, such as intrinsic qualities of the learners.

Deterministic versus stochastic models

The question is whether the limited range of behaviors that might respond to demography are actually relevant to the archaeology. Unfortunately, there's no way to predict which behaviors ought to respond to demography in this way. You might find a really clever way to test the hypothesis, even without knowing -- that was one of the features of Henrich's (2004) paper that first presented the model. I think in the current case, we can start here: If the authors' model were true, then demography would exert a deterministic effect on technology. A larger population would have a higher average "skill" level, which (by the authors' model) would allow the development of more complex culture.

When it comes to individual artifacts, demography's effect is stochastic. The development of technology has been path-dependent, with different populations following different paths. Sometimes those paths have included similar features, sometimes not. The same idea that spreads in some populations may fail to spread in others, despite the same demographic conditions.

For example, the Aurignacian split-based bone point is an intrinsically unlikely artifact. Most people in the world did not produce them, even though bone points were fairly common, especially in groups who used small-projectiles. Carved ivory figurines, on the other hand, are not nearly so unlikely; many peoples in the world have produced them. But some populations did so at very low population sizes and densities, while others have made carved ivory figurines only after reaching very large population sizes with highly specialized division of labor. Large populations make it more likely that we'll see carved ivory figurines, among other things, but they do not determine that such figurines will be present. In other words, population size is one factor affecting the stochastic appearance of these artifacts.

OK, but what if we try to generalize beyond individual artifacts or traditions and consider "modern human behavior" as a whole? Isn't there some general and abstract factor that might change deterministically with demography? To test that hypothesis, we need to (a) develop some accurate measure of the abstract factor, and (b) observe it to be deterministically influenced by demography.

Here's an example: For our work on the acceleration of recent adaptive evolution, our hypothesis was that a deterministic model based on recent demographic expansion could describe the number of new selected mutations in human populations. We tested the hypothesis by developing a measure for selection, and by showing that the numbers of variants matched the predictions of the deterministic model. This global conclusion about the number of variants holds despite the fact that any particular case of selection on a gene depends on many stochastic factors, including the occurrence of a favorable mutation, its escape from genetic drift when rare, and the function of the gene relative to recent human ecological changes. In the limit of large numbers, these random processes do not obscure the deterministic effect of population size.

Now, for archaeological observations, we could in principle follow the same procedure. If there is an abstract factor of "modern behavior", we might develop an accurate measure of it by understanding the relationship of the abstract factor and particular artifact types. That's the reason why archaeologists have devoted such extensive effort to defining "modern human behavior." The entire goal of defining "modern human behavior" is to make archaeology an instrument for measuring the cognitive advancement of prehistoric groups.

Yes, there's some irony here. Many archaeologists don't want to "invoke" cognitive capacity, even as they define "modern behavior" as a proxy for it. Artifacts certainly change stochastically. If we wanted to test a stochastic model of change, we might as well use artifacts directly. But that might not allow us to test whether the demographic factor was more important than other factors, such as developmental or ecological ones. Can we expect some combination of artifacts to behave deterministically?

The current paper chooses a simple threshold definition for the abstract factor: the Blombos incised ochre artifacts and pierced shells define the same level of "modern behavior" as the early Aurignacian of Europe. Why those two populations? Why those two behaviors? Why ignore much earlier engraved lines from other places, or pierced artifacts made by Neandertals? The paper doesn't make any serious effort to defend this measure of an abstract factor underlying "modern behavior".

I think at a minimum, the authors need to show that their measure of "modern behavior" is replicable and predictive outside the context of these two populations. If engraved lines can be a threshold measure of "skill", then they should reliably appear in some contexts and not others. If pierced shells can stand in for other elements of behavior, like small game exploitation or projectile use, then show the strength of the correlation. If they can't stand in reliably for their abstract factor, then they need to find some combination of observations that can. If there is no combination of observations that proves reliable, then their model cannot validly apply.

The second necessary element for testing the deterministic model is to show whether the measure is deterministically affected by demography. On this score, the paper is much more convincing: Their demographic model cannot explain the distribution of their measure of "modernity".

Oh, I know, the conclusion of the paper says the opposite. But look at the data: The model predicts that southern Asia should have Upper Paleolithic-like industries beginning long before they appeared in Europe, and that southern Africa should have retained Upper Paleolithic-like behaviors throughout the last 90,000 years or more. Neither of those predictions holds up. The authors don't consider the mtDNA evidence for population growth in the New World (where art and ornamentation are rare among Paleoindians) or Australia (which underwent substantial complexification during the Holocene). The comparison of Europe and South Africa is an assumption of their measure, not a prediction or conclusion.

The model really only gets one prediction correct: The West Asian record undergoes an Upper Paleolithic transition at around the same time as Europe. And even on that score, one may quibble: was the Levantine initial Upper Paleolithic earlier than Europe or later? Does the European mtDNA expansion, which mainly consists of mtDNA lineages derived from West Asia, record European demography or West Asian demography?

They're left making a variety of ad hoc arguments to explain why the model doesn't fit the demography: maybe the mtDNA samples don't represent Late Pleisocene populations exactly; maybe the population really shrank in post-Howieson's Poort South Africa even though the mtDNA (and a lot of archaeology) say it didn't; maybe there were recurrent bottlenecks and expansions not covered by the mtDNA demographic models. When ad hoc hypotheses add up so quickly, there's often much more parsimonious option: maybe the model is wrong.