Watson / Jeopardy 'robot', whats so cool?

At first we have to define that Jeopardy is a gambling game like roulette or horse-betting. The chance to find the correct answer is

1:100000 (100000 is the number of words in English language). According to pure statistics, the chance to win is very very small. That Watson did win can only be explained by cheating. Yes, Watson cheats the game to be better than randomness. He played unfair to its opponents in using additional information to better than pure randomness. It is the same strategy which is called "watching inside the roulette-table" (Kesselgucken on german).
Reply to
Manuel Rodriguez
Loading thread data ...

At the beginning of the first show, it was explained that Watson was receiving the questions as "text files". No voice recognition, no OCR, no awareness at all of what was happening in the game (other than knowing the state of the board). It was, as far as I could tell, just being a question answering machine with a little hard coded logic for picking questions and making wagers.

It certainly was not being told what the other contestants guessed, because when given a chance to guess after Ken missed an answer, Watson gave the exact same answer - the one that was flashed up at the bottom of the TV screen to the viewers as Watson's best answer even before Ken gave his answer.

That sounds reasonable.

I just pulled up the video on the youtube.

The category name was "ALSO ON YOUR COMPUTER KEYS". No mention of the word keyboard.

I suspect this is an example of how real world interaction gives humans a semantic grounding which is missing in a program like wanton that is forced to try and find all meaning in language documents alone.

It's very hard to understand what a keyboard is, or what "ON your keys" means, if you don't have the ability to see and interact with things like keyboards.

Keyboards are very common objects to humans because they are now a ubiquitous part of our lives. The more exposure we have to anything, the higher it is ranked in our web of associations. Watson, like Google, is forced to create the probability webs based on how often words are used in association with other words in these language documents. How often do we write about our keyboards? Nothing like the amount of time we spend using them.

The instant we read "on our computer keys" what pops into our heads? The lables that show up on our keyboard.

What if you have never seen, or used a keyboard, and all you knew, were the millions of wikipedia articles you had read? What was "computer key" a reference to in wikipedia? An encryption key? Lets say it made the association correctly to they "keyboard" article in Wikipedia. It's full of 1000's of words, and the labels on the keyboard, are in a jpg image that Watson probably couldn't even "see". So where as we instantly read that clue and think of the printed labels on our keys, Watson at best "thinks of" a few very long and boring wikipedia articles and has no concepts of what the labels on the keyboard actually are (because he's never seen or used one and doesn't have the level of understanding to know what it means for a human to use one.

Not having the 3D spacial temporal sensory knowledge to ground language to puts any machine that "understands" only by a sensory grounding to text documents at a severe disadvantage whenever it comes to talking about things that exist in the real world. I think the ubiquitous wording of "on your computer keys" were too abstract to allow Watson to find the connect when they were asking about a piece of woman's clothing called a "shift". The threads of connection in the associative web were just too weak (if they were even there at all).

There is however, a Shift article in wikipedia which lists Shift for computer keyboard and Shift for clothing, so the hints did exist in wikipedia. I bet if the clue used the word Keyboard instead of key, it would have gotten it.

Wikipedia is an obvious thing to use for such a task. It would be the first document I would turn to to feed data to a program like Watson for playing Jeopardy.

Makes sense. What is "key the shit out of his car?". :)

Based on what google returns it probably had a good link to "keyboard" from "computer keys". But what it probably failed badly at was understanding "on your computer keys" was a link to the very limited set of words and symbols printed on the keys. You can't read the article on keyboard and get any good sense from those words alone (with no semantic grounding to a real keyboard), what is printed on the keys. Just imaging if you read the same page in a foreign language you didn't understand, and tried to figure out what words in that document, might link to other documents related to clothing and find the answer as "shift". The word "slash" for example isn't in the keyboard page. Though the word "shift" is there many times.

I think the way Watson works is actuality very close to the way the human brain works. I think that likewise, once we see a Nova show that explains how the bran works (after it's first figured out of course), you will likewise see it "as a trick". All technology looks like magic until you understand it.

At their core, Watson and Google are doing the same sort of thing. They are taking a set of inputs that define a context (keyword list for Google, Category and question for Watson), and they are searching a probability web defined from past experience to fill in the missing pieces (such as the word "keyboard" was one of the missing pieces of context from the clue "on your computer keys". Watson searches the probability web for an answer - normally one word but sometimes a phrase - where as Google searches for web pages. They both map an input, to an output by using a large probability web of associations.

I think the brain is doing the same thing, though its input is much more complex, that is, a continual stream of sensory data, and it's output is more complex - a continual stream of output actions. The mapping it must do is likely very similar - it maps the inputs though a probability web to produce what is the best output "answer" based on the associations defined by the input context.

All intelligent human behavior, including how we think, and how we use memory, and do planing, I think can be reduced down to a simple input to output mapping problem though a probability web of associations that are conditioned first by experience (what we are exposed to) and second by reinforcement learning (what the brain is hard wired to be reward for - preventing pain, getting food).

So though at the high level, human behavior is beyond complex, at the low level, I think it's driven by a machine that really is not doing anything all that different, than what Google and Watson are doing. And when we fully understand it, it will seem like "gee, that's not what I was expecting - it seemed more magical than that".

So I would say, don't underestimate what has been achieved with the likes of Google and Watson.

Reply to
Curt Welch

Reply to
John Fields

PolyTech Forum website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.