February 2011 Archives

One step up from the AstroCube we have on Daden Prime in Second Life, we now have a whole sim running locally in OpenSim dedicated to stellar data. We are bringing data in through LibOMV so we are plotting anywhere on the sim at around 10 points per second directly from a CSV database.

For nearby stars the database we are using is the excellent one at http://www.projectrho.com/smap06.html. This has almost 45000 entries for stars out to around 5000parsecs (although 40000 are within 300parsecs), and is based on the ESA Hipparcos Catalog, the Yale Bright Star Catalog (5th Edition), and the Gliese Catalog of Nearby Stars (3rd Edition) with some extra data thrown in.

We can plot out to 128m from the centre point, ie whole sim, and the images in this post are based on plotting all stars within 20 parsecs (1 parsec is about 3 light years), around 1900 stars, at a scale of 5m per parsec!. Stars are coloured and sized based on spectral type (OBAFGKM) and Luminosity (I to VI).

The backdrop to the plot is a 256 x 256 cube with a texture of small white stars for effect purposes only. In the image above you can also just make out my avatar floating by the Sun.

We are not working on building the interactive features so you can interrogate each star and plot additional information (like exoplanets).

Zooming in towards Sol. The bright white stars are Sirius (lower right) and Procyon (upper right).

Sol. Proxima/Alpha Centauri group just under my arm!

With planetary discs switched on. Discs are colour coded by planet type, red/orange for Jupiter type, blue for Neptunes, green for terrestrial. About 100 planetary discs within range.

The upshot of all of this of course is not only to show that we can plot larger and larger quantaties of stellar data (the eventual aim is to get the 40000 stars within 300 ly plotted), but that we can use OpenSim to plot almost any sort of data in 3D.

They say it's not the winning that counts but the taking part. Watson managed to do both -- but perhaps the winning actually counted for less. All the indications are that Watson's win was as much about his - or its -- speed on the buzzer as for his knowledge.

As competitor Brad Rutter said, "On any given night nearly all the contestants know nearly all the answers, so it's just a matter of who masters buzzer rhythm the best."

So the impressive thing is that Watson had sufficient knowledge to be a credible contestant, the winning was secondary.

(Image courtesy IBM.)

While of course the sheer power of the computers involved and the man-decades of development count for a lot, the real key to Watson's success was in its use of semantic markup. Most chatbots (of which Watson is a specialized type) rely on matching an input pattern against a database of possible pattens, and then returning the response programmed for that pattern. This is how Pandorabots (much used in Second Life) work, and to an extent our own Discourse system. The problem is that this approach does not scale, and the bot "knows" nothing -- it just matches characters.

With a semantic approach the bot has a web of knowledge. The most pure way is to use triples, very simple relationships between an object ("sea"), an association ("has colour"), and a subject ("blue"). When these elements are then linked into an ontology of things (blue is a colour, sea is made of water, sea is a habitat, sea is a geographic feature) then the bot can begin to make deductions and links between the words in the input text.

This means he can deduce the things that the question might be alluding to, and try and map that to an answer. Watson is not the first chatbot to use semantic markup (we've also been doing similar things since 2008 but just not with IBM's resources -- see our Halo bot talking about ants and stars), but we are in no doubt that semantic representation will be key to chatbots in the future.

As to whether Watson is the future and represents a great leap for artificial intelligence, it all depends on what you are looking for. If you see AI as being about god-like bots, Terminator's Skynet, or, more prosaically, expert systems, then yes this is a step forward. Wolfram Alpha was one step towards the "global answer engine" (but Google and Wikipedia more so), and Watson is another step. But these are brute-force solutions to brute-force problems.

One of the truisms of machine intelligence is that as long as a machine is unable do a particular task, people say that the task requires intelligence. But once the machine can do it, then they say that the task didn't need intelligence after all. Chess is a good example. Once seen as almost the best measure of an "intelligent" person, after Deep Blue beat Kasparov in 1997 chess was seen as more of a brute-force problem, and no one claimed that Deep Blue (or the chess app on your smartphone) was intelligent.

The issue is that Watson is an "answer engine," not a "conversation engine," and definitely not an "artificial general intelligence."

Trying to mimic the flow of human conversation is something that still appears to be beyond our reach, and for this Watson is actually heading in the wrong direction. Human conversation is often very imprecise, even inaccurate, but it flows. If I asked Watson what 16,353 times 9,543 was he'd probably know -- but most humans wouldn't. In creating more "human" bots we actually find we have to try and dumb the computer down.

For example, in a virtual world we don't want our bot to say "you are at 123.2, 31.42, 78.12," but rather "you are standing on my foot," or "you're a couple of feet from the door." While giving IBM credit for their achievement with Watson, I'll personally be more impressed when IBM create a believable four-year-old rather than a Jeopardy champion.

These two paths, towards a global answer engine, and towards an analogue of a "real" human, are both valid, and to an extent connected, but they are different paths. And I feel that humantity will be more challenged ultimately by the possibilities of the latter than by Watson and his descendants.

This column by Daden MD David Burden originally appeared at Hypergrid Business, and is reprinted here with permission.

One of our ongoing interests (helped along by a current research contract) is in how we use virtual worlds as immersive data visualisation spaces. You may be familiar with our Datascape virtual war-room, but our current focus is on plotting large amounts of data over  a sim-sized area. The challenges here are a) to get enough data points in and plotted in a reasonable amount of time and b) work out how best to represent them, c) how best to interact with them and d) how to navigate through the dataset.

We started using our existing SL plot-bot to plot objects in OpenSim - the downsides of that were plot range (you have to move the bot with 10m of each plot point) and getting data in (notecards limits you, HTTP is too slow/cumbersome.) But by doing a pre-sort of the data we could get the bot to plot about a data point a second over the whole sim and brought about 1000 data points in. 

Another approach was to use a programme to write an XML object file which Hippo could then import into SL as a single object. That gave us a nice 1000 object cube (see below), but was again a bit cumbersome, particularly is you wanted alignment and positioning with the sim itself.

We then started looking at RegionModules in OpenSim, as they appear to be th sensible solution, but it looks like they've been borked for several releases, and we'll have to wait for 0.7.1 for them to work again.

So the latest work has been using LibOMV to control a bot and get the bot to build the points. We were very pleasantly surprised to find that the bot can rez an object anywhere in the sim, no distance limit, so reading a data file and plotting points became almost easy. A shot from a test run is shown above, the bot plotting points every 0.5m across the sim at a rate of about 10 points per second. Admittedly it only got to just over 4000 points before it gave up, but that was on my laptop. Over the next weeks we hope to get to reliable plotting of 10,000+ data points. We'll keep you posted.

By the way the terrain file is an outline of the UK we brought in a while ago - combining a RAW file like that with some 2D overlay textures and the data plotting and we begin to have the beginning of quite a nice system.

Two weeks and two trips to London Olympia, leaving the office behind to unleash my wrath of learning and technology experience upon the world! :)

Last week I attended the Learning and Skills / Learning technologies show in Olympia, jam packed with over 40 international speakers and over 4,000 attendees. A really exciting and interesting event with a lot less glitz and gimmicks as previously experienced at BETT the week before. Apart from an original shiny PAC-MAN table-top arcade on one of the exhibition stands, and in my book providing a very sophisticated element of geekdom class.

Many exciting international exhibitors including Intellego, Pearson, Skillsoft and Assima, all showing off the latest in e-learning and training products. Most of which looked really intuitive and well designed incorporating the last in Web 2.0 technology. I found this a refreshing look into the world of corporate e-learning and certainly more valuable than some of the products available to schools at BETT the previous week.

However virtual worlds hardly got a look in. The only company sporting the "virtual flag" for training was skills2learn and yes I must say their products looked good and well executed, The only pitfall I could see was it seemed to be a solo experience, with no community / social element with other users. Which I feel is a massive benefit for the use of this type of technology - Imagine a learner struggling on a question and being able to seek support from peers from within the platform.

http://www.skills2learn.com/virtual-reality-simulation.html