Saturday, May 27, 2017

Reading up on biogeography part 4: track analysis for bioregionalisation

With two papers left, I was wondering whether there would still be any point to going on. The last two use track analysis and area cladograms, respectively, and those were already used by the first and second paper, so would there be any new insights into the methodology of pan- and vicariance biogeography?

However, the next paper,

Martinez et al., 2017. Biogeographical relationships and new regionalisation of high-altitude grasslands and woodlands of the central Pampean Ranges (Argentina), based on vascular plants and vertebrates. Australian Systematic Botany 29: 473-488.

... uses track analysis at least partly to do something different than the previous instance. There is the question of the "relationship" of a biome, but then there is also bioregionalisation. So that is a new angle.

The idea seems to be relatively simple. As before, the panbiogeographer looks at the occurrences of species, draws minimum-distance lines ("tracks") between them, and then identifies areas where the tracks of several species overlap as "generalised tracks". In the present case, a very short generalised track is then "used to recognise natural areas in terms of their biota because they result from more or less consistent overlapping distributions of two or more endemic taxa".

Okay, same question as always: does this make sense?

Well, more than the claim that generalised tracks are always evidence of vicariance, which this paper kind of only makes in passing (while, weirdly, explaining the panbiogeographic reasoning in words so identical to those used in the Romano et al paper that I wonder if they were in both cases copy-pasted from Croizat). To me the approach just seems part unnecessarily complicated, part not data-rich enough.

As for the first, yes, an area with several endemic taxa may well deserve recognition as a natural unit, a vegetation zone, a biome (whatever) in some area classification. But if the idea is to identify areas defined by endemic species, why do we need a track analysis as an intermediate step? Why not simply plot the occurrences of endemic species? At that point all the information is there, and tracks, generalised or not, do not add anything.

As for the second, as I mentioned before there are several other methods available for bioregionalisation. Some use clustering approaches to group grid cells or other small areas into larger areas based on shared species content or even the relatedness of those species. The newest ones use modularity or map equation analyses to examine networks of species and the grid cells they occur in; in contrast to clustering, where it is the researcher's somewhat subjective choice how many clusters to accept, these network approaches have algorithms for deciding more objectively how many truly distinct units there are.

In other words, in my eyes track analysis seems to be superfluous to requirements if we are merely interested in the simple measure of shared endemics, and it is unable to provide the depth of information that could be obtained from examining other shared distribution patterns.

Sunday, May 21, 2017

Reading up on biogeography part 3: Hopping between islands yes, hopping from continent to island no?

The third vicariance biogeography / panbiogeography paper in the special issue is

Grehan JR, 2017. Biogeographic relationships between Macaronesia and the Americas. Australian Systematic Botany 29: 447-472.

Despite being very long, its gist is easily summarised:

The mainstream explanation for the occurrence of plants and animals on the Macaronesian islands (Canary Islands, Madeira, etc.) is that they must have got there via long-distance dispersal, often from Africa but sometimes from the Americas, because the islands are of relatively young volcanic origin and distant from other land masses. However, the "model-based approaches" that this conclusion is based on cannot be accepted because they supposedly assume dispersal and ignore the possibility of vicariance.

This is followed by many pages of example cases of plants and animals illustrated with maps and phylogenies. It is not clear to me what that is supposed to show, because without a time axis it doesn't move the inference either way; at best it could show that some of the groups have a pattern that is consistent with vicariance, but if a lineage is too young then vicariance is still out, and the same if the lineage is much older than the island.

Finally, there is some speculation, again illustrated with maps, about whether there were always volcanic islands in the same area, all through from the time when the Atlantic started to open. They would have been transient on a geological scale, so the local lineages supposedly produced by vicariance when Africa and the Americas started moving apart would have had to island-hop as new volcanoes rose and older ones eroded away, over more than 100 million years.

In contrast to the previous two papers I did not really gain new insights into the methodologies favoured by vicariance biogeographers. In a sense the present paper is closer to an opinion piece or perhaps a review article than to a research study.

The supposed assumptions of "model-based approaches"

The paper claims
"Model-based approaches to Maccaronesian biogeography assume the that the [sic] sequence of phylogenetic relationships reflects a sequence of chance dispersal. Although often cited as Hennig's progression rule, it is not a rule but an assumption that does not address the equal applicability of sequential differentiation across a widespread ancestor."
And further on:
"Model-based methods use chance dispersal to explain divergence and allopatry, ..."
Unfortunately this claim at least is demonstrably false. There are various models available to do ancestral area inference (see this graphic as an example), and DIVA and very popular DEC, for example, include vicariance. That's what the V in the acronym DIVA means! If a model-based analysis with a model that allows vicariance infers no vicariance then we can assume it is not because the model does not allow vicariance, but because the data didn't support that conclusion.

I am also reasonably certain that Hennig's progression rule does not only apply to long distance ("chance") dispersal but would just as well apply to a series of range expansions followed by speciation events across a single land mass. It simply applies the principle of parsimony to historical biogeography, arguing that if several lineages along a grade occur in an area then that would probably, all else being equal, have been (at least part of) the ancestral range, because other explanations require more dispersal and/or extinction events.

It is interesting, by the way, how the word "model" seems to be used in this context, as if a mathematical description of a system is something bad.

What distribution patterns would we expect under vicariance and long-distance dispersal, respectively?

"The progression rule also assumes that a 'basal' grade is located in the source region or centre of origin, but some Macaronesian clades are basal to large continental clades, and there are also clades with 'reciprocal monophyly' in which a diverse Macaronesian clade is the sister group to a diverse continental clade. These phylogenetic and geographic incongruities do not arise in a vicariance interpretation of phylogeny, because a basal clade or grade marks only the location of the intial phylogenetic break or breaks within a widespread ancestral range."
I don't really understand the reasoning here. The idea seems to be that if an island clade is nested within a continental grade, then it may make sense to conclude dispersal, but if an island clade and a continental clade are sister to each other then it is somehow "incongruent" (with what?) and can only be explained by vicariance. Why?

I would look at the nearest outgroup to get more information, but even if that occurred in neither region then we would still have to ask if additional continental or island lineages may have simply gone extinct. The key questions are whether the lineage split is so recent that it happened considerably after continental break-up and whether an island lineage is older than the island(s). Really I don't see how we can conclude anything with confidence without a time axis.

Perhaps the idea is to equate "distribution of the species along a basal grade is evidence of a centre of origin" with "absence of such a basal grade is evidence of absence of a centre of origin"? If so, that would not be logical; absence of evidence for A is not evidence for not-A.

The paper also discusses other patterns, in this case non-overlapping ranges of related species (allopatry):
"Model-based methods use chance dispersal to explain divergence and allopatry, and yet allopatric divergence requires isolation, which cannot exist if there is effective dispersal."
The point of the second half of this sentence is a false dichotomy set up between dispersal that is so frequent that it makes speciation impossible and no dispersal at all. It seems obvious to me that the excluded middle is dispersal that happens but is too rare to make speciation impossible.
"In the same way that allopatric lineages within Tarentola are incongruent with the expectations of chance dispersal, so too is the allopatry of Tarentola and its New World sister group."
Again this makes no sense to me whatsoever, and again there seems to be some very black-and-white reasoning behind it: if species can disperse to distant islands everything should occur everywhere; but we observe that all species do not occur everywhere, so we have to conclude that dispersal is completely impossible. But this is one-to-one equivalent to the argument that you cannot produce random numbers with a die because when you cast it the second time it came up with a different number than the first time. Really, that seems to be the logic here.

One might also add that there is another fairly obvious reason why one would find patterns of allopatry even if the same region was reached two or three times by the same lineage: competitive exclusion. It is a well established, empirically tested insight of biogeography that islands (and by extension restricted areas in general) have a carrying capacity, both in overall diversity and in the number of species trying to occupy about the same ecological space. In the case of islands in particular, their species diversity is a function of size (the more land, the more species, mostly because lower area increases extinction rate) and distance from the nearest larger land mass (the closer, the more species, mostly because of higher immigration / dispersal rates filling up the species pool).

This makes a lot of intuitive sense. Assume you have a seed of a continental shrub species blown onto an island that so far has only been colonised by mosses, lichen, one species of grass, and a bunch of insects eating the former. Your shrub niche is still free, and there is nothing on the island that is adapted to eating you, so even if at first you are in a bit of trouble genetically (inbreeding) and ecologically (not used to this soil and climate) you have a reasonable chance of establishing. Now fast forward 500,000 years, and the single seed of that shrub has diversified into six species occupying every niche on the island that they could adapt to in that time, forming thick scrubland from coastal dunes to the highest peak. A new seed of a related continental shrub species ends up on the island - but now everything is occupied by relatives that have become well-adapted to this new environment. Are we really surprised that the second comer will have a harder time establishing?

Time-calibration of phylogenies, again

We had that one already in the Ung et al paper, but once more:
"Model-based methods, with rare exceptions, present molecular divergence ages as falsifications of early origins, at or before continental breakup, even though they are calibrated by fossils that can generate only minimal divergence dates. Although it is widely claimed that molecular-clock analyses are generate [sic] evidence of dispersal (Sanmartin et al., 2008), molecular divergence estimates artifically constrain the maximum age of taxa that may be much older than their oldest fossil or the age of the current island they occupy (Heads 2009a, 2012, 2014a, 2014b, 2016)."
I like the little caveat "with rare exceptions", although it is unclear what it refers to. But it is not a method, but the researcher using a method, who would draw the conclusion that a lineage diverging 12 Mya would not have diverged because of a tectonic event that happened 120 Mya. And yes, that conclusion makes a lot of sense to me, and no, "model-based" methods do not magically transform minimum ages into maximum ages. This has been discussed repeatedly in rebuttals to Heads' papers. What is more, people have run analyses using the alternative approach suggested by Heads and in the present paper and found that the results are generally absurd, such as pushing the age of the daisy family back before the origin of multi-cellular life.
"The timing of ancestral differentiation may be assessed either by fossils (including molecular extrapolations) or tectonic-biogeographic correlation."
First, fossil calibration or using estimated substitution rates are really two completely different data sources, so the former does not really "include" the latter. Second, using continental breakup to calibrate splits in the phylogeny would, as mentioned before, be circular reasoning. It would build the assumption of vicariance into the analysis to subsequently conclude vicariance as a result. I think that's not how science is supposed to work.
"Fossil data provide only the minimum known-age of taxa and [sic] fossils are often lacking for clades of interest to Macaronesia. In tectonic correlation, the estimate of clade age is more precise, because it refers to a particular, dated event, rather than a minimal (fossil-calibrated) age."
Yes, a fossil provides a minimum age. But unless I severely misunderstand something, a continental break-up could, at best, provide only a maximum age, if we assume that divergence would not have been possible before break-up. (And even that seems fishy to me, given that there are plenty of speciation events on the same landmass.) If it were to be taken as "precise" that would, once more, automatically exclude the possibility that the divergence happened later, after dispersal from one continent to the other, and that would be circular reasoning.

Even the vicariance approach would need long distance dispersal to work

Finally, I am puzzled by the idea of how the lineages would have stayed in place after the supposed vicariance event that would have happened long before the present islands came into existence:
"Island biota survives erosion and subsidence of island habitats by local dispersal onto newer volcanoes"
What I don't get is this: if a vicariance biogeographer can accept that a species hops across the ocean from one volcanic island to another, why can they not accept that it hops across the ocean from Africa onto one of the volcanic islands? What's the difference? Why is this discussion taking place again? I must be missing something very subtle here.

Friday, May 19, 2017

Reading up on biogeography part 2: Panbiogeographic Track Analysis

The second paper in this little series of posts is
Romano MG et al, 2017. Track analysis of agaricoid fungi of the Patagonian forests. Australian Systematic Botany 29: 440-446.
What I appreciated about reading it was first that it was concisely written, and second that it gave me insight into the Panbiogeographic methodology of Track Analysis. It had so far been merely a bunch of arcane terms to me, which of course makes it impossible to judge its meaning. And in contrast to the previous paper, which left out most the details of its methodology and instead referenced earlier papers, this one gives a clear explanation. This kind of stuff is exactly why I am reading through the journal issue.

So, how exactly does Track Analysis work?

First, you need species with disjunct areas of distribution - or at least species that are poorly enough sampled that they appear to be disjunct. Then you draw a line along the shortest distance between any two of their occurrences. Let's assume we have a species occurring on two islands of this little landscape I just generated in GIMP:


Panbiogeographers call this red line, with the occurrences of the species forming the end points, a Track.

If you have more than one species showing the same Track, you promote that line on the map to a Generalised Track:


To cite the present paper, in panbiogeographic logic "a generalised track ... allows inference of the existence of an ancestral biota widely distributed and fragmented by vicariance events, suggesting a shared history."

Now you may come up with other tracks in the same study group that do not run parallel. Where generalised tracks cross each other, panbiogeographers draw a circle with an X in it and call that place a Node, like this:


In this case, their interpretation is that this is "a complex area, where different ancestral biotic and geological fragments interrelate in space-time as a consequence of terrain collision, docking or suturing".

Aaaaand... that was it, really. Draw some lines on the map, conclude vicariance and "complexity". The rest of the conclusions in the present paper are largely about the need for more sampling, and that fungi can also be used as a study group.

Does this approach make sense?

Unfortunately, I don't really see it. The logic behind the panbiogeographic interpretation of Generalised Tracks is that patterns of disjunction shared by several taxa are evidence of vicariance, presumably because they assume that chance dispersal would have to be utterly random and create different distributional patterns in each and every species.

But a little contemplation should blow that idea out of the water. There are several other good reasons why disjunct ranges can be shared across taxa. One would be an a priori lack of alternative habitat - if you have two wet patches and otherwise only steppe, then all wetland species will be restricted to those two patches, even if one of the two wetlands was colonised from the other entirely through long distance dispersal. And that restriction alone will produce a shared history, without vicariance. Another option would be prevailing wind or ocean currents, which make long distance dispersal decidedly more probable in some directions even as it is still a stochastic process (dice, but a bit loaded) and, more importantly, not vicariance.

The interpretation of Nodes as showing things like terrain collision also seems to be missing a few crucial steps, at least in my eyes. Don't get me wrong, I am as aware of fossil ranges being an important part of evidence in geology as the next biologist, but still I'd actually prefer to consult a geologist instead of trying to deduce geological history from patterns of distribution alone.

Finally, this whole approach appears to have a weakness that seems quite critical. Science does not proceed by knowing how to confirm, it proceeds by knowing how to reject a hypothesis. Now the question here is this. Yes, panbiogeographic track analysis is apparently designed to conclude vicariance and an area being "complex". But if a disjunction really has not been caused by vicariance, how would a panbiogeographer conclude that? Would they ever do so?

That, alas, is left unexplained, at least in this paper.

Sunday, May 14, 2017

Reading up on biogeography part 1: area cladogram for the southwest Pacific

As mentioned in the previous post, I am hoping to learn more about the reasoning behind panbiogeography and area cladistics by going through the relevant papers in the recent special issue of Australian Systematic Botany. Starting with area cladistics, I have today finished reading

Ung V, Michaux B, RAB Leschen, 2017. A comprehensive vicariant model for Southwest Pacific biotas. Australian Systematic Botany 29: 424-439.

To the best of my understanding the key steps of the study can be summarised in a very bare-bones fashion as follows. The authors...
  1. State that very little is still known about area relationships, as most research focuses on ancestral area inference for individual taxa.
  2. Summarize at length - over nearly four pages - the geological and tectonic history of the region. I cannot judge any of this at all and will consequently take it as given, although it is puzzling that no reference seems to be provided for the claim that the now largely submerged region had much more dry land when it broke away from Gondwana.
  3. Divide the study region into areas - the details don't matter for present purposes.
  4. Compile 76 phylogenies for plant and animal taxa occurring in the study region, and replace the species with the combination of areas in which they occur.
  5. Discuss the 'problems' of incongruence between the area relationships in these individual phylogenies, of terminal taxa occurring in several areas, which they call "taxonomic paralogy", and of the same area occurring in different branches of a phylogeny, which results in what they call "paralogous nodes". They decide to exclude these confounding nodes and to use only "paralogy-free subtrees" by applying a "transparent method" that I had not heard of before.
  6. Turn the trees into "three-item statements" and use those to produce a consensus area cladogram.
  7. Present the consensus area cladogram.
  8. Argue that one larger area that they had hypothesised is not a "real biogeographic entity" because it is paraphyletic on the area cladogram.
  9. Argue that New Caledonia's "highly endemic flora and fauna are ancient" because of its "basal" position on the area cladogram. I am not sure that this follows, and am a bit concerned about the potential of scala naturae thinking here, but that is not the main point here.
  10. Agree with panbiogeographer Michael Heads that any and all time-calibrated phylogenies are unreliable. Then they proceed to a lengthy attempt at time-calibrating their area cladogram based on plate tectonics.
I would like to explore in a bit more depth items #1, #5 and #10.

Does the concept of area relationships even make sense?

I cannot say that this paper has me convinced. To quote a few sentences where the authors themselves discuss problems:
In real-world situations, individual areagrams are unlikely to be congruent with each other and the problem, therefore, arises as to how best to deal with this incongruency [sic]. The main sources of incongruency [sic] are the occurrence of widespread taxa (multiple areas on a single terminal, or MASTs, for short), redundant areas (resulting in taxonomic paralogy), missing areas and inadequate methods of analysis (dos Santos 2011). Redundancy, the repeated occurrence of the same area in different branches on the areagram is nigh on universal and results in paralogous nodes. [...] [These] yield no information about area relationships and obscure the real relationships between areas.
Honestly, when I read this I am drawn to a very different conclusion than that we have to exclude all "paralogous nodes": maybe there is so much noise because stuff moves around too much. In other words, the concept of an areagram or area cladogram makes exactly as much sense as trying to force members of the same sexually reproducing animal population into a phylogenetic tree. Where there is no phylogenetic structure, phylogenetic trees are not an appropriate representation of the data.

Another issue I wonder about is the use of the term paralogy in this context. The word comes from gene evolution. Imagine a gene has duplicated in a distantly ancestral species, and subsequently both copies A and B evolved to have different functions. (This is, of course, one of the main ways in which new genes come into existence.) All descendant species inherit both genes. If we now look at a bunch of descendant species and want to figure out their relationships, we need to make sure we compare only the A copies or only the B copies. Comparing the A copy from one descendant with the B copy of the other misleads our analysis; the A and B copies are called paralogues of each other, and the A copies from different species are called orthologues of each other.

What I do not understand is how the situation in areagrams is supposed to be equivalent enough to use the same terminology. Areas are not genes that are inherited by species lineages. At best, it is the other way around: if the assumptions of area cladistics are true (which I doubt), then species lineages are comparable to genes inherited by areas. The same mistake as taking two paralogues as orthologous in genetics would then be to treat two species lineages in different areas as orthologues although they already diverged before continental breakup.

But the way the word is used here is in the former sense, when contemplating areas on a phylogeny, not when contemplating lineages in areas. This use of genetic terminology is rather confusing, I have to say.

What is the problem with time-calibrated phylogenies?

The open access de Queiroz paper in the same issue does a good job at discussing Heads' and the present authors' criticism of molecular dating, so just very quickly, there are two arguments here:

First, that
using substitution rates derived from modern taxa and then applying them over evolutionary time, often to groups only distantly related, is not justifiable
This is true as far as it goes, but the problem is that to the best of my understanding for the conclusions favoured by Heads to be realistic, substitution rates would have to be off by an utterly unrealistic factor. We are talking cases here where he sees a divergence as having happened tens of millions of years ago when the molecular data say a few million years. And why would we assume such massive shifts conveniently in just the direction needed to make vicariance a viable explanation, and in the absence of any other argument? Sorry to say, but that looks a bit like ad-hoccery to me.

I hope this is not taken to be too inflammatory, but it reminds me of those young earth creationists who are worried about the starlight problem and then argue that a few thousand years ago the speed of light must have been orders of magnitude higher. There is, indeed, a very practical parallel: just like the creationists in question do not take into account what such a change would do to other physical parameters (E=mc^2, meaning that our planet would have been incinerated), so in this case nobody seems to consider what a massively higher mutation rate would have done to the biology of the affected species.

The second argument is that
the same can be said for dating phylogenies using the age of the oldest fossil, which, despite giving only a minimum age for divergence, becomes a maximum estimate by proxy (Heads 2014b)
As has been discussed at length in rebuttals of Heads, including again in the aforementioned de Queiroz contribution, this is half nonsense and half, let me say, odd. It is nonsense in the sense that fossils are indeed used as minimum ages, not as maximum ages. I have myself recently used the R package chronos to time-calibrate trees, and you simply tell the analysis to make a divergence no younger than so and so, and that's that. Admittedly you generally also want to have some realistic maximum age for the entire tree, but that can be way higher than any minimum age you set. In fact, I wrote a blog post about this stuff not too long ago.

In Bayesian analyses, it is true, it is necessarily the case that there will be a limit to how much older than the fossil the results can realistically be because calibration is usually done with priors. The user sets a prior probability distribution where the probability of divergence, which necessarily has to add up to 100% over all possible times, will become so close to zero as to make no difference if we only go far enough back in time. It is, after all, impossible to stretch 100% out over infinity years and still have 10% per million years left.

But here is where the argument also gets distinctively odd. What Bayesian phylogeneticists do in practice is to set a relatively high probability around the time where the fossil was dated, and then have it peter off towards the past. The question is now: what else would one do? Is it not eminently reasonable to assume that the further into the past we go from the known existence of a lineage, the less likely it is that it already existed? Surely it is reasonable to assume that if the oldest known fossil of a plant genus is from 20 Mya, then it is quite likely that the genus already existed around, say, 21 Mya, a bit less likely that it existed 30 Mya, still less likely that it existed 50 Mya, and vanishingly unlikely that it existed as long as 200 Mya?

The problem with time-calibrating a tree based on plate tectonics is, in turn, that it front-loads the analysis with the assumption that there is no dispersal between areas. For the purposes of the discussion around vicariance and dispersal it is circular reasoning.

But to end on a positive note, despite approvingly citing panbiogeographers the authors of the present paper actually do not seem to argue that dispersal between areas is impossible; they merely kick out the data that I would interpret as showing such dispersal to infer the 'real area relationships'. Admittedly that could be seen as equivalent to kicking out all the genes I share with my father to claim that my genetic relationship with my mother is the 'real' one, but well, it still makes more sense to me than hostility to the mere possibility of dispersal!

Thursday, May 11, 2017

Panbiogeography and area cladistics galore

Today the newest issue of Australian Systematic Botany came out, and oh boy is the content interesting. It is the first in a series of special issues on biogeography - but that is not the main point, to which I will come later.

I have tried before to systematise for myself what biogeography is actually about. What is its research program? Trying again, and perhaps in a way that reflects my current thinking:

1a. Inferring ancestral ranges, and closely related to that...
1b. Inferring biogeographic events and their timing.


This kind of research is focused on a given clade and tries to understand how its species came to occupy the ranges they do today. It uses a number of approaches and software tools that attempt to infer ancestral ranges given a generally time-calibrated phylogenetic tree of the study group, contemporary distributions at the tips of the tree, and a model specifying what biogeographic processes are 'allowed' to happen. Examples include originally parsimony-based Dispersal And Vicariance Analysis (DIVA), the Dispersal, Extinction and Cladogenesis model (DEC), and others.

A typical result would be on the lines of, "the ancestral range of this genus was in the south-east of the continent, and we estimate ca. three sympatric speciation events and ca. two vicariance events in its history", often illustrated with a phylogeny whose branches are labelled with the relevant ancestral ranges and biogeographic events.

2. Species distribution modelling

Research in this field tries to estimate where a species can occur, usually given presence data for the species and climatic, soil and other data for those known locations. This can be used, for example, to predict to where approximately in Australia an invasive species could spread out if it were introduced from its native range in, say, South America. Computationally intensive, species distribution modelling is a relatively recent development. That being said, it was the big hot new thing when I did my first postdoc, so recent is to be taken relative.

Obviously, a typical result would be a map with different colours indicating different probabilities of the species being able to exist in those locations.

3. Spatial studies

This field divides a study region into cells, often equal area grid cells, and attempts to quantify diversity metrices such as species richness, endemism, and phylogenetic diversity. Hotspots of diversity can then be targeted for conservation, or they simply provide information on the evolution of present diversity, especially if they are hotspots of palaeo- or neoendemism. This work has only really become possible with the availability of large biodiversity databases of geo-coded specimens.

A typical results would be on the lines of, "the study group shows the highest endemism scores in the south-west and the tropics".

4. Bioregionalisation

The idea here is to distinguish bioregions across the landscape that are significantly different from each other in their species or lineage content, and to figure out where their approximate borders are. Traditionally this was done very intuitively and based mostly on the presence or absence of key taxa. Today researchers often use computers and grid cell-based approaches similar to those in spatial studies, only that they compute pairwise dissimilarity scores between grid cells. Cells are then clustered into bioregions or, in the most novel approaches, submitted to network analysis.

A result might read: "Our analysis shows four major bioregions, the monsoonal tropics, the Eremaean, the south-west, and the temperate south-east. The border between the monsoonal tropic cluster and the Eremaean cluster is, however, considerably further south than estimated by a previous study..."

5. Area cladograms

And this is where I am leaving my comfort zone, because while I have used #1 and #3 and at least dabbled in #2 and #4, this one is weird to me and will probably remain so.

The idea in this case is to use areas or bioregions as the units of an analysis that is supposed to show how the areas are related. In other words, something like a phylogenetic analysis of areas, using their species content as data, and with a result on the lines of "the Australian temperate rainforests are sister to the New Zealand temperate rainforests, and together they are sister to the Patagonian ones" (not necessarily a true result, just to get the concept across). There are a few methods available for this, and they are generally parsimony based and by now quite dated.

The obvious problem here is that this whole procedure is based on a number of assumptions that I can only consider dubious. Just like phylogenetic reconstruction of the tree of life must assume, in that case rather sensibly I believe, that there is no significant gene flow between, say, cattle and primroses, building a tree of bioregions must assume that there is no significant dispersal or species exchange between the various area it uses as units of analysis. And that is where it all falls down for me, because of course species disperse happily from area to area. There are no barriers that are remotely as strong as as the barriers to gene flow between different species.

The present issue of Australian Systematic Botany

So we arrive at the present issue of Australian Systematic Botany, which is, as mentioned, the first in a planned series on biogeography. My personal perception is that of the above fields of research, the cutting edge is today in ancestral range inference and spatial studies. Species distribution modelling is often more seen as part of ecology rather than systematics; the scope for large numbers of bioregionalisation studies is obviously somewhat limited, given that there are considerably fewer bioregions than species; and I thought that area cladograms were more a thing of the 1980s or so.

But the papers in the present issue show that they are still being done - and so is panbiogeographic track analysis!

This is going to be very interesting, because when I read either of these approaches I have the same feeling as when examining some of the pro-paraphyly literature: intellectual challenge in the sense of having to understand a mode of thinking that is very, very alien to me. But that just makes it more important to try and follow the reasoning, even should it ultimately not be found convincing.

In particular I am looking forward to seeing a track analysis in action when I come to those papers, because so far it really has not clicked for me what they are supposed to show and how their conclusions can possibly be justified.

To summarise, the articles in the issue are:

1.&2. Two very short introductions.

3. Alan de Queiroz rebutting an earlier article by panbiogeographer Michael Heads. This one is open access, and otherwise stands out in that it seems to be the only article by a mainstream biogeographer. As I pretty much agree with everything it says I will not have any comments on it.

4. Ung et al. constructing an area cladogram for "southwest Pacific biotas", with the abstract indeed containing phrases such as "the islands of the Southwest Pacific are more closely related to each other than they are to Australia". Interestingly, they call their results a "model".

5. Romano et al's panbiogeographic track analysis of agaricoid fungi of the Patagonian forests.

6. An extremely long article by panbiogeographer John Grehan on relationships between America and Maccaronesia.

7. Martinez et al. conducting a panbiogeographic track analysis on plants and animals of the Argentinean pampas.

8. And with Corral-Rosas & Morrone another area-cladistic analysis, this time with Mexico as the study area.

Ancestral range reconstruction for individual clades or spatial analyses, on the other hand, are clearly MIA. So at a minimum one would have to say that this is, at the moment, still a rather narrow representation of the field of biogeography.

Friday, May 5, 2017

A good read on superhuman artificial intelligence

This essay written by Kevin Kelly must be the most sensible text on superhuman artificial intelligence (AI) and the allegedly imminent "singularity" that I have ever read.

Although it appears to get a bit defensive towards the end, I am in complete agreement with all main points. In my own words, and in no particular order, I would like to stress:

There is no evidence that AI research is even starting to show the kind of exponential progress that would be required for an "intelligence explosion".

There is no evidence that intelligence can be increased infinitely; in fact there are good reasons to assume that there are limits to such complexity. What is more, there will be trade-offs. To be superb in one area, an AI will have to be worse at something else, just like the fastest animal cannot at the same time be the most heavily armoured. Finally, we don't want a general purpose AI that could be called "superhuman" anyway, even if it were physically possible. We want the cripplingly over-specialised ones. That is what we are already doing today.

Minds are most likely substrate-dependent. I do not necessarily agree with those who argue that consciousness is possible only in an animal wetware-brain (not least because I am not sure that the concept of consciousness is well defined), but it seems reasonable to assume that an electronic computer would by necessity think differently than a human.

As for mind-uploading or high-speed brain simulation, Kelly points out something that I had not previously thought of myself, even when participating in relevant discussions. Simulations are caught in a trade-off between being fast because they leave lots of details out on one side, and being closer to reality but slower, because more factors have to be simulated. The point is, the only way to get the simulation of, say, a brain to be truly 1:1 correct is to simulate every little detail; but then - and this is the irony - the simulation must be slower and more inefficient than the real thing.

Now one of the first commenters under the piece asked how that can be true when emulators can simulate, 1:1, the operating system of computers from the 1980s, and obviously run the same programs much faster in that little sandbox. I think the error here is to think of the mind as a piece of software that can be copied, when really the mind is the process of the brain operating. Simulating all the molecules of the brain with 1:1 precision, and faster, on a system that consists of equivalent molecules following the same physical laws seems logically impossible.

Finally, one point that Kelly did not make concerns the idea that a superhuman AI could solve all our problems. He discussed that more than just fast or clever thinking is needed to make progress, experiments for example, and those cannot be sped up very much. But what I would like to add is that of our seemingly intractable problems the really important and global ones are political in nature. We already know the solutions, it is just that most people don't like them, so they don't get implemented. Superhuman AI would merely restate the blatantly obvious solutions that human scientists came up with in the 1980s or so, e.g. "reduce your resource consumption to sustainable levels" or perhaps "get the world population below three billion people and keep it there". And then what?