Watson and the Future of Everything: Continued

My posting has been scarce as of late. I've been involved with a project I wanted to finish as quickly as possible. Now I need to try to get back into a blogging rhythm. It's not always easy.

Last post was several days ago. (I'm being kind to myself.) It was Watson and the Future of Everything, in which I waxed poetic about the significance of IBM's Watson computer performing well on Jeopardy. I have seen various people express considerably less enthusiasm that I did, those people being concerned about the impact on jobs, and dissing on the Watson because it doesn't really understand, it just computes.

Both categories of Luddites are correct. Watson-like computers will enable people to me more productive in their work, and therefore more valuable in their work, and therefore better paid. Productivity generates wealth. Inefficiency breeds want.

And true enough: Watson doesn't really understand anything. It merely computes, sometimes badly. In that sense it is quite human.

IBM, on their website, offers three videos on three areas of application for Watson-like computers.

Healthcare: "Medical records, texts, journals and research documents are all written in natural language – a language that computers traditionally struggle to understand. A system that instantly delivers a single, precise answer from these documents could transform the healthcare industry."

Finance: "Enormous amounts of data are generated every day in the financial industry. Watson, the IBM computer system designed to compete on Jeopardy!, has the deep analytics capability that could help businesses extract knowledge from this data in order to identify patterns and make more informed financial decisions."

Customer Service: "IBM experts share their thoughts on how DeepQA technology could help transform the customer service industry into a faster, more accurate experience."

Wow!  Way to take something really exciting and make it boring. But IBM is a business and naturally focuses on those applications that can provide substantial return for all the money, time, and talent they have invested in Watson. When I was coming to know Watson, however, I was thinking differently. I was thinking that a Watson type computer could have kept Byron Case and Michael Ledford and a whole lot of other innocent people from ever being convicted.

Assume we have a computer that understands natural language and is designed to estimate probabilities of innocence. We would constantly feed it the latest findings in pathology, serology, ballistics, fingerprinting, fire science, and other forensic sciences. We would provide provide it with common sense insight into life, insight such as time cannot run backwards, people cannot be in more than one place at the same time, consistent stories are more likely to be correct than ever changing stories, snitches are usually less reliable than priests, confessions that violate physical laws are probably false.

Even before arrest, all parties involved in a case can feed Watson case information, ask Watson to provide a probability of guilt and/or innocence, and ask Watson to substantiate its assessment.

In the case of Byron Case, Watson could read the police report that said the victim was found with her eyes open at 3:47 AM, and the ME report that reported her corneas were clear. Watson could compare that  information against arcane knowledge (unknown to anybody involved in the case at the time) that corneas cloud over in less than three hours after death if the eyes remain open. Watson would realize that the victim died some time around midnight, that the state's primary witness claimed she actually witnessed the murder near sunset, and conclude that the state's primary witness was unreliable. Watson would realize further, from reading the police reports, that Byron had an unassailable alibi for the actual time of the murder. The probability of Byron's  innocence would be near 100%. The probability of his guilt would correspondingly be near 0%. Assuming the State was acting in good faith, they would not even charge Byron, much less convict him and sentence him to life without parole.

In the case of Michael Ledford, Watson could read descriptions of the fire scene photos and learn that the circuit breaker panels shows evidence of smoke damage within the panel but not around the panel. Watson could also read that those smoke patterns show some of the breakers popped to the closed position during the fire and some did not. Watson could read conflicting theories on how the fire started (either a candle tossed into a living room chair or a short in a living room light circuit.) Looking through the world's knowledge, it would find no correlation between candles thrown into chairs in a living room and simultaneous internal heating in a circuit breaker box in a bedroom. It would find, however, that on occasion circuit breakers fail to pop when the circuit shorts, and that causes them to overheat and smoke. That results in evidence of a fire both near the short and in the circuit breaker box. Even though the Michael Ledford had confessed (then quickly recanted), the evidence of fire at the circuit breaker box would prove the fire to be electrical rather than incendiary.

In the words of a great philosopher, you may say I'm a dreamer, but I'm not the only one. At least I hope not. Months ago, I wrote to Google about the possibility of coupling their computing power with their legal scholar database to help identify many of the quarter million people I calculate we have incarcerated today. I'm still waiting for an answer.

Late last week, I wrote to IBM about the possibility of using a Watson like machine to improve the quality of our judicial system. I did not propose removing the human element. I proposed only that we improve the accuracy of the system by increasing the quantity of information and quality of the analysis.

I heard back from IBM the next morning. The person who responded described the concept as "very interesting." Perhaps he was only being polite, but I'll take it. He explained he would pass the idea along to his supervisor who would reply in further detail. I now await that reply. I actually expect to hear something further from IBM, though not necessarily the answer I would like to hear.

My scoring right now is IBM 1, Google 0.

If I cannot interest IBM, I will try Microsoft. If I cannot interest Microsoft, I will try Wolfram Alpha (a mathematically based search engine.) Then I will try something else.

Our wrongful conviction problem is so serious, we need something on the order of a Watson computer coupled with a public acknowledgment of the problem to make even a dent in the number of people already (or soon to be) wrongfully incarcerated.

Stay tuned.

Watson and the Future of Everything

The world changed Monday night. Perhaps you missed it. Perhaps you were eating dinner or watching the news. Perhaps you were taking a walk or making love. Perhaps you were watching Jeopardy and did not fully appreciate the world changed before your very eyes.

On Monday, the two best Jeopardy players in the world squared off against IBM's Watson computer, and the world changed.

Let me begin by introducing the contestants.

On your far left is Ken Jennings. Ken won 74 Jeopardy matches in a row, and earned $2,520,700 in the process. He also earned $2,000 in his 75th game when he came in second to Nancy Zerg, who came in third during her next game. Jennings then won $500,000 when he placed second in the Jeopardy Ultimate Tournament of Champions. That's a grand total of $3,022,700.

It is almost statistically impossible that Ken Jennings won that many games in a row. Prior to Jennings' streak of 74 games in a row, the longest streak was a pitiful (only by comparison) 8 games in a row by Tom Walsh. Jennings streak was more than 9 times longer than Walsh's streak. By comparison, Joe DiMaggio's 56 game hitting streak, considered by many to be the greatest feat ever in baseball, is only 1.3 times as long as Pete Rose's 44 game streak, second longest in the modern era.

On your far right is Brad Rutter. Brad Rutter has never lost a Jeopardy game to another human. When Rutter first appeared on Jeopardy in 2000, contestants were limited to five games in a row. (That rule was eliminated in 2003, one year before Ken Jennings began his remarkable run.) Rutter, however, later won each of the three Tournaments of Champions in which he was invited to play.

In the third of those tournaments, the Ultimate Tournament of Champions, Rutter beat Ken Jennings and took home $2,100,000 for his efforts. In all, Rutter has earned $3,255,102 for his initial five Jeopardy games and his three Jeopardy tournaments, making him the top Jeopardy earner of all time.

In the center is Watson, a massively parallel IBM supercomputer.

Born in 2007, Watson has (at the hands of some amazing teachers) learned to understand the subtleties of natural language. Consider Watson's rapid fire questions to some of Jeopardy's ambiguous, nuanced questions in the practice session recorded before the formal contest.

Since Jeopardy might present any of an infinite number of convoluted answers to the contestants, there was no way Watson's teachers could simply program it to select the right question from a database of question/answer pairs. They had to teach Watson to understand natural language. That was the challenge. That's why they created Watson. It was to be the first computer to understand the intricacies and subtleties of natural language well enough to compete on Jeopardy.

Watson's knowledge is kept in much the same way as Google stores all its data. When faced with a Jeopardy answer, Watson uses keywords, much as Google does, to collect a long list of possible answers from its stored information.

It is at this point, however, that IBM has gone beyond Google and put itself back into the lead as the world's foremost computing company. Watson is programmed to use its natural language skills to decide which of the many possible answers is most likely correct, and to answer only if it is confident beyond some threshold value.

On its web site, IBM provides a 22 minute video explaining how sophisticated Watson's decision-making can be. Consider the Jeopardy answer "In May 1898 Portugal celebrated the 400th anniversary of this explorer's arrival in India." The graphic below represents one possible result Watson might quickly find. Based on a keyword search, Watson finds a match for "celebrated", "may", "anniversary", "Portugal", and "India". Watson also finds that the link is associated with some guy named Gary.

As it turns out, "Who is Gary?" is unlikely to be the question Jeopardy was searching for.

The link segment reads "In May, Gary arrived in India after he celebrated his anniversary in Portugal."  As humans, we can easily see Gary is not who we are looking for, because we understand natural language. No computer, other than Watson, shares that understanding.

For Google, the Gary link would be a terrific match. For Watson, however, it would be but one of many to be rejected as improbable. Instead, Watson would home in on a Vasco de Gama link, which reads "On the 27th of May in 1498, Vasco de Gama landed in Kappad Beach."

The astonishing point here is that the Vasco de Gama link has only one matching keyword, and that keyword is the commonly used and ambiguous word "may." Holy linguistics, Batman! There are millions and millions of documents out there with the word "may" in them. Why did Watson decide to consider this one?

Watson picked up on this link because Watson has been programmed, as have we humans, to apply temporal reasoning, geospatial reasoning, and statistical paraphrasing.

While the Vasco de Gama link doesn't actually mention 1898 as being the 400th anniversary of anything, it does mention the year 1498. Watson knows that 1898 minus 1498 is equal to 400, and  therefore recognizes two very strong clues in the link, strong clues missing from the Gary link.

Furthermore, while the Vasco de Gama link doesn't actually mention India, Watson knows that Kappad Beach is in India. It knows that from a few of the other millions of links it has in its memory. So that's another big clue.

Finally, Watson understands statistical paraphrasing. It knows that "arrival in" is a good match of "landed in."

Watson combines all the information at hand and assigns a high probability that Vasco de Gama is the correct result. No other answer it considers has a higher probability. Since the Vasco de Gama probability is greater than the confidence threshold, it buzzes in and says, in the best computerized voice to date, "Who is Vasco de Gama?"

When Watson finds a reasonably probable answer, it typically buzzes in (using a mechanical finger) more quickly than its human competitors. This too is impressive, given all the mental work it has to do. I learn from the video that if Watson used my home computer as its processor, it would need several hours to decide each question. Watson, however, uses thousands of processors working at the same time, working in parallel.

So how did all this pan out in the actual game? As I said in the beginning, the world changed on Monday night. Watson finished the first game (played over two nights) with $35,734. Rutter had $10,400. Jennings had $4,800.

After three days and two games, the totals were:

Watson: $77,147

Jennings: $24,000

Rutter: $21,600

In my next post, I'll discuss what this might mean to the quarter million people we have wrongfully incarcerated in America.