How Natural are Trees?

Hello, I followed the discussion on some Chandler lists for quite some time, and here are some things I wanted to mention that 'accumulated' over time, so this became a little bit lenghty. Sorry.

As a start, I want to use the discussion here http://wiki.osafoundation.org/twiki/bin/view/Chandler/BrowserDesign which I found very interesting and right to the point. But I would like to direct your attention to some work in the cognitive sciences that I consider highly relevant and groundbreaking for a project like chandler. This is not meant to be criticism, but an attempt to gain new metaphors, paradigms etc. for what Chandler could/shoud be from my point of view. It's also not a 'final treatment', but more like a RFC in the proper sense of the word.

How 'natural' are trees?

Following the discussion of cognition scientist George Lakoff in his seminal 'Women, Fire, And Dangerous Things - What Categories Reveal About the Mind', the tendency to put things that belong together in the same basket may be 'natural', but not for the reasons that western thought has assumed up to now.

The 'classical' view of categorization goes back to Aristotle and supposes that there is always some necessary and sufficient condition which is used to decide if something belongs to a category or not. For someone to be a woman, it is necessary but not sufficient to be a human being; sufficient is for her to be female and of a certain age. This is the classical, 'philosophical' point of view which, of course, expresses how things *should* be (if philosophers ruled the way we think).

If we look at the way people actually make use of categories, a different picture emerges: From empirical psychological research we know that people use *prototypes* to build categories, that is, by deciding if something is sufficiently similar to some 'best case' or 'best example', we decide that the thing at hand 'is a bird' or 'is not a bird'. Then there is the case of Wittgensteins 'family resemblances', which means that the items in a category may share features from a common set of features, but there is not necessarily a subset of features that they all share.

This is relevant for a PIM since outside the narrow world of bureaucracies (where everything has an ID) or commerce (where everything has an ID and a price), a lot of categories we use are built on these 'weak' conditions (family resemblance, prototypes) and have fuzzy boundaries. (In my oversimplifying world view, iTunes belongs to commerce, of course - every track has an 'ID' and a price. But iTunes will not be of any help if you want to track which of Miles Davis' compositions and themes appear on his live recordings from the early until mid-seventies, and where. He would usually make his band play a new theme every now and then during the performance, and the track lengths and titles on the albums more often hide than reveal this.)

So in the trees vs. semi-lattice debate, the tree's shortcoming is not only that we have necessarily to decide for one view at a time. The problem goes deeper. Whatever we want to achieve, it turns out that in several critical phases of the process, we have to deal with categories that rest on weak conditions like "x is relevant for y" or "x could be interesting as a possible solution for y if our main interest is z", where x is the set of items satisfying the condition. The 'container' logic of classical categories might very well keep us from seeing the relationsships that are *really* interesting for us. (Coming back to the Miles Davis example above, my ideal iTunes interface would require MP3 ID-Tags that support additional information on the musical content at any point on the track's timeline. Which is still a relatively 'treey' requirement, but demonstrates how quick we come to the necessity to expand the tree logic in ways that are hard to satisfy if the software is built on just one type of 'atoms', in this case, the tracks with their features.)

Trees are just one kind of cognitive model

But Lakoff's argument goes further. In his view, a treelike taxonomy is just an application of a very common, very basic cognitive model (which is derived, ultimately, from our experience as upright, spacially oriented beings in a three-dimensional, yet ever-changing space we move in), the model of the container. For any given structure of containers in containers, it's not possible for something to be at the same time inside and outside a container at any given level. Either the item satisfies the required condition to be inside a specific container, or it doesn't. Unfortunately, this is a 'topos' where Aristotle, the medieval world view where everything has a place in the divine hierarchy, binary logic and the structure of our hard drives collaborate to keep up one very special, very restricted 'cognitive model' of the world and the things 'in' it: the tree, or 'containers in containers'.

My impression is that we shouldn't try to get rid of the semi-lattice all too quickly. While I second the criticisms of Plumbdesigns visual thesaurus and its blindness towards the relevance/irrelevance or meaning/meaninglessness of the visual cues it gives, I'm sceptical if the proposed 'stepwise trees' = lattice can overcome the problems with trees outlined above.

Let's assume I want to use my PIM to write a book on Miles Davis' electric period (it has already been written: "Miles beyond" by Paul Tingen. Great book.). So I put everything I have in my database: Events, persons, track listings, records, re-issues, studio sessions, comments by critics, comments by people who knew Miles, interviews, related musicians (musicians who played with him, musicians who influenced him, musicians who were influenced by him, musicians who tried not to play like him, musicians who were in some respects like him like Hendrix or Prince), some texts on racism and other social circumstances, on the 50s, on flower power, other fields of creativity he touched (painting), etc. - the problem is: There are no trees! All these items, quotations, dates etc. are related to Miles Davis ("Blues is relevant for his tone") in one way or another, but apart from decades there are no self-evident trees (and if there are, like the category of 'style', I would probably try to destroy them to make my view of Miles valid).

Relations other than 'container in container'

I probably would like to work with very un-treelike relations like 'similarity' (in which respect?), 'relevance' (for what?), 'influence', 'met with', 'disgusted'. Of course, all the people who talk about Miles Davis in my book form a group and can be considered a 'tree' which is only one step deep, but that's just a matter of convention. I could as well call them a heap.

And yet, for a software to enable me to work more efficiently with all that information, all these items need to have 'markers' or 'cues' that may indicate, however vaguely, where they belong. E.g, for the basic event of Miles Davis talking to one of his musicians about Music, ideally I have a (rough) date, the place and the topic (Music). Filtering all conversations by time, person and topic, my PIM could come up with a lattice of all (known) occasions where Miles talked to one of his musicians in the sixties, including the one where he approached Herbie Hancock to say: »Don't play the butter notes.«

OK, as I write, I see that this is probably what you are heading for with "tree+tree+tree=lattice", but I wonder if it might be better to come up with a weaker data structure than a tree and see the *relation* as 'fundamental', e.g. the relation 'talks to' which would be a function like (talksto "Miles Davis" "Herbie Hancock" topic:"Music" date:1969) and as such a 'knowledge atom', a proposition that would enable my PIM to answer the question "Did Miles Davis ever talk to Herbie Hancock?" "Yes, several times."

Conceptual Blending

Another author you should probably look into besides Alexander and Lakoff is Edwin Hutchins who wrote a seminal book in cognitive science about the distributed cognition of ship navigation, "Cognition in the Wild". Lately, he refined some of his findings to make the connections with Turner/Fauconnier's 'conceptual blending' approach more evident, talking about 'material anchors for conceptual blending'.

So what's this about? According to Mark Turner and Gilles Fauconnier, 'conceputal blending' (cf. The Way We Think, 2002) is a fundamental operation of 'cognitively modern human beings', that is during the recent approx. 50.000 years. In conceptual blending (cb), two conceptual input spaces are 'blended' to form a new, 'blended' space that allows for new cognitive operations that weren't possible in either input space ('emergence'). This may or may not include the use of external symbols or symbol systems. 'Space' is of course a metaphor here, we might as well talk about 'domains' or 'realms'.

Hutchins is mainly interested in the cases where 'material anchors' are used to do this kind of 'blending', as is the case with certain instruments in ship navigation that 'embody' earlier acquired knowledge about positions, the stars, tides etc. - the 'embodied knowledge' is blended with information on the current situation to make predictions like the exact time of the next flood exactly here, etc.

For computer interfaces (which are by themselves 'conceptual blends', at least the visual ones), one of the most striking example of conceptual blending is the spreadsheet program 'visiCalc' (a forerunner of Excel) that was programmed late in the 1970. It's available somewhere on the net, its size in the DOS version is 28k.

The power of visiCalc (that became a killer app for the Apple) comes from the blend it allows to run: The traditional table as a collection of lists is blended with something that is only possible with a digital computing device: instant evaluation of numeric expressions. We have to remember that at the time when viciCalc was written, the usual way to use a computer was the command line, so in the best case, an expression was evaluated in the next line (more often, however, much further down). The idea to have a 'living' table that completely changed whenever the user changed one of the input values was nothing short of revolutionary and very intriguing. It became possible to play around with different scenarios with the least possible effort (just change the input value(s) that could be interesting) and see the results immediately.

I curious if you think we could see Chandler as a 'material anchor' for conceptual blends that include even more input spaces (the Semantic Web, RSS feeds, web services) and in a much broader fashion than in the case of a spreadsheet. It's a fascinating idea to see all of my data sources as one input space, my interests (my project, the book I write, my weblog) as another input space and the Chandler browser UI as the blending of those two possible. It's important that it's semi-automatic: As soon as something *relevant* changes in one of the input spaces, the consequences will have to be visible in the UI.

Although I know that it's hard to compare numbers on a spreadsheet with natural language or XML on the net, but I don't think the analogy is too far-fetched to be productive. According to Gregory Bateson, an information is a difference that makes a difference, and this is what I expect the interface of my PIM to reflect.

best regards, Christoph