Wednesday, May 31, 2006 |
|
|
Microsoft today announced a new version of VS "for database professionals." It
seems that the emphasis is on maintaining database integrity while
evolving the data model: a worthy first goal. Perhaps as significant as
the release itself is the way in which the SKU has been developed: the
team was only formed in September '05, promises to deliver by the end
of this year, and "agile techniques" are supposedly part of the
process. And who's in the midst of
all this excitement, but my old friend Thomas Murphy, who was the Product Review Editor of Computer Language back in the early '90s.
|
|
|
|
|
Monday, May 29, 2006 |
|
|
An Old Neural Net Programmer on Jeff Hawkins' Model of Intelligence
Monday, May 29, 2006
7:07 AM
There's one big issue with the simplest possible 3-layer artificial neural network:
Created with Microsoft Office OneNote 2007 (Beta)
One place for all your information
Download: An Old Neural Net Programmer on.one
|
|
|
|
|
Sunday, May 28, 2006 |
|
|
Sunday, May 28, 2006
9:28 PM
Daniel Crenna, one of the finalists in the "Made in Express" contest, felt I went too far in dismissing the entrants when I said their projects were unrealistically ambitious. One of his co-finalists is a professor of robotics, Daniel is confident of his approach, etc. Okay. Why someone capable of writing, in a month or two, a realtime 3-D vision system in C# from scratch is looking to win a $10,000 prize is beyond me, but bully for them for doing it.
Crenna is developing a domain-specific (visual) language for poker robots. He says that he doesn't intend to "advance the state of the art (not in this competition, anyway), but I will do my best to make what is currently available more accessible," with a drag-and-drop interface. This is a worthy goal and not in the same realm of ambition that the 3-D vision system is. I think that it still lies in the realm of "if you can do this, you shouldn't be giving it up for a $10K prize," but that's his business, not mine.
Modeling poker is a fascinating problem. I have just subscribed to Crenna's blog and look forward to reading on his progress: I hope he'll forgive me kibbutzing.
The thing about Poker, and Texas Hold-Em in particular, and Tournament Texas Hold-Em in double-particular, is that it brings the forefront the problem of modeling intentionality. First-order intentionality is when you look at your cards and say "I believe I have a strong hand," (and therefore, I will play). That's easy. The great thing about Texas Hold-Em is that while there's variability in what cards will come up, the variability in what cards you have is very small and the importance of first-order intentionality is minimal.
A "poker intelligence" based on first-order intentionality would have a table of starting hands that are "likely to win" and bet on, say, A-7 or better, fold anything else. After that, it would be driven by pot odds. It would ignore other players betting patterns and would be very easy to beat.
Second-order intentionality is (he played and, therefore,) "I think he has a strong hand," and would be necessary for any non-trivial poker intelligence. So, for instance, if the other player opened, the poker intelligence might "put him" on an A-7 or better and compare the pot odds against various predictions of what the other player might do. Obviously, people play differently, but you should get some results if you had a parameterized model ("Aggressive player," "Passive player," etc.).
Poker betting signals third-order intentionality: (I bet aggressively out of position so) "I think he thinks I have a strong hand." And even the lamest poker player understands bluffing "I think (if I bluff) he will think that I think I have a strong hand." To call a bluff requires a decision about fourth-order intentionality: "He thinks I think he thinks I have a good hand," and, just to take things to what is generally considered the human limit, tournament texas hold 'em happens so fast that you have to model your opponent's model for dealing with bluffing: fifth-order intentionality.
By the time you get to fifth-order intentionality, you're verging on comic territory -- "Only a fool would put poison in the cup in front of him!" I don't think that fifth-order intentionality is necessary for a non-trivial poker intelligence, but I do assert that third-order intentionality is necessary, since that's the level at which bluffing takes place.
Another possibility is to collapse the model into statistics: model your opponent as "10% of the time, he's betting over his card's true odds (aka bluffing)," but putting such a parameter to an opponent's play is very difficult since it is difficult to get enough data about your opponent's real situation versus the evolving pot odds. Again, especially in tournament hold 'em.
Created with Microsoft Office OneNote 2007 (Beta)
One place for all your information
Download: AI for Poker.one
|
Sunday, May 28, 2006 10:19:17 PM (Hawaiian Standard Time, UTC-10:00) | Disqus link | AI | DotNet
|
|
|
|
|
Unpredictability and recognition systems
Sunday, May 28, 2006
12:18 PM
In reading Jeff Hawkins book On Intelligence I came upon this great anecdote about developing Graffiti:
"I recognized that people were willing to learn a difficult task (typing) because it was a reliable and fast way to enter text into a machine. Therefore if we could create a new method of entering text with a stylus that was fast and reliable, people would use it even though it required learning. So I designed an alphabet that would reliably translate what you write into computer text; we called it Graffiti. With traditional handwriting recognition systems, when the computer misinterprets your writing you don't know why. But the Graffiti system always produces a correct letter unless you make a mistake in writing. Our brains hate unpredictability, which is why people hate traditional handwriting recognition systems." (Emphasis added)
To this day, I prefer Graffiti for PDA input, although I would love Shark/Shapewriter (which bolsters Hawkins' point even further). On the other hand, I prefer the TabletPC's TIP and correction UI to Graffiti; I'm not sure it's faster, but the correction UI is good enough that using it is predictable. Voice recognition systems, though, definitely produce the "unpredictable == hateful" reaction in me.
Created with Microsoft Office OneNote 2007 (Beta)
One place for all your information
Download: Unpredictability and recognition.one
|
|
|
|
|
|
Sunday, May 28, 2006
9:11 AM
On the suggestion of John Lam, I bought Jeff Hawkins' book On Intelligence and pretty much read it through in one sitting. That was possible because it's a very accessible book and also because, to a large extent, I was already exposed and sympathetic to the grand themes: minds are what brains do; pattern recognition, association, and prediction is close, if not indistinguishable, from "intelligence;"and the brain's mechanisms for doing such things are based on highly interconnected hierarchies of localized neuronal structures that exhibit "fire together, wire together," reinforcement / learning. (For those interested in such things, it should be noted early that Hawkins punts on the problems of qualia, so as far as I am aware, that debate still ends with Dennett and Searle.)
Hawkins biggest hypothesis, though, is one that I find intriguing but far from self-evident: that a single algorithm that produces "predictive ability [from] a hierarchical memory," is sufficient to achieve intelligence. Hawkins was a founder of Palm and Handspring and I was surprised that On Intelligence is rested on the structure of the neocortex and not on computational structures. Of course, a premise is that the neural structures he talks about are computational, and he presents his theory in terms of a few block diagrams, but the book is several steps away from presenting the source code, as it were.
This is a little disappointing, because Hawkins' hypothesis could be tested with relatively easy computational experiments. That the human brain has a gazillion neurons and a bazillion interconnections is largely irrelevant if the phenomena that Hawkins claims arise from relatively small collections of neurons. A testbed that maps well into Hawkins theories is the world of "complete information, zero-sum, binary placement gridded games." Tic-Tac-Toe, Reverise/Othello, and Go are all this type of game: they all take place on a grid and involve sequential placement of opposing symbols. Unlike Poker, there's no hidden knowledge or probability issues; winning is binary; and unlike Chess or Checkers, you don't have to understand movement. A "play" is the changing of a position from an indeterminate state to a determinate one and the consequences of that play. Helpfully, Tic-Tac-Toe is close to trivial and Go is computationally more difficult than Chess. Any model that could solve Tic-Tac-Toe and simple be scaled to Go would be an incredible triumph.
And although we probably don't have much insight into what Hawkins' theories predict for the form of a Go-solving intelligence, they are testable with Tic-Tac-Toe. It should be straightforward to create a "cortextlike memory system" whose "sensory inputs" are Tic-Tac-Toe placement sequences. So, for instance, a blank grid followed by an X in the upper-left and then an O in the lower-right might be encoded as: "??? ??? ???, X?? ??? ???, X?? ??? ??O". If Hawkins' model is correct, then one would expect certain things to be emergent in the self-organizing higher-levels of his hierarchies. Note that we are not testing the ability of a connectionist system to play Tic-Tac-Toe (Tic-Tac-Toe is possible to "solve" with a traditional neural net approach, Go is almost certainly not). Specifically, it follows from Hawkins' theories that in Tic-Tac-Toe, if the first play is to a corner then a single high-level component should be responsible for "expecting" play to the opposite corner -- no matter which corner is initially played -- even if the system has never been exposed to play from a particular corner. Optimal play in Tic-Tac-Toe is very straightforward and recognizable, rotational equivalence is trivial compared to the types of sensory interpolation the book asserts are explainable, and yet, it is not self-evident that Hawkins' model will pass such a test.
The book has a companion site at OnIntelligence.org -- I just posted a similar comment to the above paragraph, we'll have to see if anything comes of it.
Created with Microsoft Office OneNote 2007 (Beta)
One place for all your information
Download: On On Intelligence.one
|
Sunday, May 28, 2006 2:28:40 PM (Hawaiian Standard Time, UTC-10:00) | Disqus link | AI
|
|
|
|
Saturday, May 27, 2006 |
|
|
Editing and State Transitions
Saturday, May 27, 2006
8:58 AM
I absolutely love this research paper from Microsoft that describes a UI for a recording application. The great thing is that "all" they did was rethink the classic recorder interface of "rewind, play/record, fast forward" with a new state transition for "re-record that last thought." The UI tells the story:
Isn't that great? The results are significant, too: content creators uniformly preferred the interface and, even more significantly, 84% of listeners preferred freeform recordings made with this interface to recordings made with a traditional UI. I love that delivered quality depends on editing interface -- something I've noticed when trying to write high-quality software for the Tablet PC.
I've actually thought of similar (though not as elegant) approaches to pen-written passages: accumulate short sections of expressive thought and assemble them using an interface different than the cursor-based WYSIWYG of a Word / classic keyboard editor.
Created with Microsoft Office OneNote 2007 (Beta)
One place for all your information
Download: Editing and State Transitions.one
|
|
|
|
|
Thursday, May 25, 2006 |
|
|
Sweet. The simple trick is to use OneNote's "send email" capablity
Created with Microsoft Office OneNote 2007 (Beta)
One place for all your information
Download: One Note Blogged from.one
|
|
|
|
|
|
For me, the VS2005 "Extensibility Project"|"Visual Studio Add-In" wizard doesn't produce the expected result -- the add-in doesn't register itself with VS. I've tried this on two machines and it fails on both. Registering an add-in is done via an XML file with a .AddIn extension that (documentation indicates) should be placed in \Documents and Settings\<User>\My Documents\Visual Studio 2005\Addins and that's where the Wizard copies the file.
Well, on my machines, files in that directory are not picked up by VS2005. However, copying the file to \Documents and Settings\All Users\Application Data\Microsoft\MSEnvShared\Addins works like a charm: debug works, the whole bit.
|
|
|
|
|
|
When you start the beta versions of Outlook or OneNote, you get a dialog box that says "Blah blah blah functionality requires Microsoft Desktop Search. Download now?" That seems dangerously close to using the Office monopoly to "tie" Microsoft's offering in personal search. Not only that, you don't even have an option of "Don't tell me again." It's either "Download now" or "Remind me later."
Personally, I use X1 for personal search and don't want to change. But for the sake of OneNote, I'd do anything. To add insult to injury, though, MS Desktop Search doesn't install on test machine! I just get an "access denied" dialog and the install quits, even when run in Safe Mode. |
|
|
|
|
Wednesday, May 24, 2006 |
|
|
The WMPlayer10SDK does not contain the expected primary interop assembly (PIA) so that Windows Media Player can be programmed from .NET languages such as Visual Basic and C#. The WMP 9 SDK puts a PIA in the /redist directory. However, the WMP10 SDK can be programmed in a managed language: you can just copy the files AxInterop.WMPLib.dll and Interop.WMPLib.dll from the /samples directory. Or, you can still download the WMP9SDK and use the old PIA.
Okay, hopefully this post now contains enough keywords to be found via search engines... |
|
|
|
|
Tuesday, May 23, 2006 |
|
|
Oh well. Being able to blog from OneNote is a dream. I had a whole bunch of posts using graphics and ink all set to go. |
|
|
|
|
Monday, May 22, 2006 |
|
|
I get 'em all for free: this is one I pay for out of my own pocket. |
|
|
|
|
Wednesday, May 17, 2006 |
|
|
I bought a dartboard a month ago to give me something to fidget with while thinking. Rather than struggle with cricket or any "real" dart game, I've been training myself with a game I call "golf." I go around the numbers in the board, and if I hit the target number in 3 throws, that's "par." If I hit two targets in 3 throws, that's "a birdie," 3 targets with 3 is an "eagle," and if I miss in 3 throws, that's a "bogey" (yes, this is asymmetrical. I said I was a beginner, didn't I?) Today I went around the board and made "par." Hurray for me.
We will now return to our regular programming... |
Wednesday, May 17, 2006 9:56:53 AM (Hawaiian Standard Time, UTC-10:00) | Disqus link | Offtopic
|
|
|
|
|
Recently, I've seen a huge jump in "Undeliverable mail" messages coming through my mail server. I've double-checked that I'm not relaying and checked a couple return path IPs, so I know that it's just jerks spoofing the "Reply-To" header, but it makes me sick to have my domains in any way associated with this crap.
|
Wednesday, May 17, 2006 9:42:48 AM (Hawaiian Standard Time, UTC-10:00) | Disqus link | Offtopic
|
|
|
|
Tuesday, May 16, 2006 |
|
|
Apparently, the “code navigation bar” in VS2005 is a
performance drain. I never rely on it, so I’ll be implementing this
change on my own system. Via Steve Pietrek
|
Tuesday, May 16, 2006 11:06:34 AM (Hawaiian Standard Time, UTC-10:00) | Disqus link |
|
|
|
|
|
In the dot-com days, my work led to three, count 'em, three patent applications. In all cases, I'd signed over the IP rights to the company. Both companies went bankrupt before the patents went through, leaving me to understand that the IP is now the property of creditors, who were free to auction it off or use it to create their own chimerical success story or, most likely, throw the CD in the bottom-drawer and wait for it to oxidize.
Anyway, I'm kind of jealous that I don't have a patent. My brother has a great one -- you know when they use a cotton swab on your laptop at the airport? Thank Steve O'Brien.
So I've been thinking that, in my copious spare time, I should get myself a patent. Sure, if they go through on the first pass (doubtful), it'll cost around $1,000 in filing fees. But, hey, which would give me more pleasure, a patent or two-thirds of a cheap HDTV? Of course, I could spend thousands of dollars on a patent attorney, but while I'm sure there are good patent attorneys that are worth every penny, I strongly suspect that a cheap patent attorney is like a cheap tax-preparer: you still do the same amount of work explaining and checking as you'd do if you file yourself. (That was certainly my experience previously.)
The nice thing about patents is that the full text of granted patents are available. It's not that hard to parse out the specialized syntax and, while it's unquestionable that I could misuse a term of art, the patent office has given patents to combovers ; is it really likely to ding me on my sloppiness with the word "whereby"?
Coming up with patentable ideas is easy. (30-second pause.) Automatic ranking of dart-throwing skill by means of a neural net. A lava-escape device consisting of a highly heat-resistant snowboard. A means of swinging from a substantially horizontal tree branch. (Oops)
Of course, I'll probably never spend the time to actually write a patent, but if I do, I'll blog the process here.
|
|
|
|
|
|
Microsoft is running a contest called "Made in Express" that will pay out $10,000 to the best application written in Microsoft's entry-level version of Visual Studio. The proposals chosen as finalists are, to put it kindly, ambitious: an AI psychotherapist (assuming that the person doesn't just cut-and-paste Eliza); an autonomous robot capable of traversing rough terrain; a poker bot simulator with drag-and-drop AI. There are others that are more feasible (the one I like is the "absurd comparisons" calculator: how many hummingbirds weigh as much as a blue whale?).
Lately, I've been thinking about how people learn programming, and it's very interesting to note how these programmers (who know enough about programming to find and enter the contest) have, let's say "unrealistic" understandings of AI and machine sensing. If you think about it, it's not absurd for someone to say "Hey, I hear Tablet PCs understand handwriting, I hear speech recognition is almost there, therefore, how hard could it be to include '3D stereo vision' in my contest entry?" Except, sadly, it is absurd.
Once upon a time, I did AI "technology transfer" work: basically, translating LISP-based algorithms into (primarily) C and C++-based algorithms (it's one reason I believe Sapir-Whorf applies to programming languages). I haven't in a long time, but I try to track development and I have a fair understanding of progress in the field. We are so friggin' far from a computational model of mental states that it's laughable (alternately, it's humbling).
AI is decades behind genetics in terms of a model. At least we know how DNA codes. And, until recently, we believed that there was something approaching a 1:1 correspondence between genes and functions. Until we starting sequencing genomes a few years ago and discovered that there are many, many fewer genes in a genome than there are traits in an organism. Even with DNA-based genetics, which is an incredible natural computational abstraction, it turns out that reality is complex. Even chaotic in the sense of depending on environmental starting conditions. (The idea that "well, the genome didn't explain as much as we thought, but the proteome (our protein complement) will," is equally hubristic and short-sighted.)
And just as we can assemble certain helpful drugs based on biologics, even with our incredibly limited understanding of how those biologics came to be, we can compute certain helpful functions (handwriting, speech) with a vastly more limited understanding how we do it natively. But at least with DNA we have a (limited) model of how things move from small to large. With cognition and perception, we have nothing even close, just very, very general theories.
For instance, there seems to be widespread agreement that "minds are what brains do." Further, I think most researchers would probably more-or-less agree that mental states are the work of specialized, individually non-accessible subsystems working in parallel, whose complex (and chaotic) interactions construct higher-level systems. I hope that's enough for that guy to create the Web-based psychotherapist.
Then we have this mechanistic understanding of neurons which, just like DNA, have these features which can be modeled computationally. And a lot of fascinating research is coming from functional imaging, so we're beginning to understand the physical structures associated with mental states. But, I guarantee you, the way that mental states arise from neurons will turn out to be far, far more messy and complex and contingent and filled with feedback loops and dependent on environmental conditions than the way living organisms arise from DNA.
And don't even get me started on consciousness. |
|
| | |