IBM Makes A $1-Billion Bet To Make Watson A Business
For the past two years, since its Watson question-answer system went on the Jeopardy! game show and beat the two best humans at this game, an IBM team with a few hundred people centered in Austin, Texas, has been working on ways to commercialize Watson and put it to work. With the launch of the Watson Group inside of IBM today in New York, Watson is entering its third phase, where IBM is investing $1 billion over several years to extend the technologies underpinning Watson and get them into mainstream products and services.
|IBM’s Watson Group headquarters in New York|
At the launch event in New York, IBM CEO Ginni Rometty said that the information technology industry was entering its third phase, and reminded everyone that Big Blue was instrumental in the two prior phases. The first phase was tabulation, which is where IBM has its heritage back in the late 19th century. This was merely counting things. The next phase, in the middle of the 20th century, was for programmable computers, which could run IF-THEN routines and perform useful calculations and manipulation of data. The third phase, Rometty said, is the cognitive era.
“It was a new species, if I can call it that,” Rometty said of the original Watson machine that won at Jeopardy! back in early 2011. “By design, it learns by experience and it learns from interaction, and by design, it gets smarter over time and gives better judgments over time. We took Watson on because it is built for a world of big data.”
The Watson system was a cluster of IBM Power 750 servers running Linux with an in-memory database. This database was fed by an Apache Hadoop system and also employed another set of Apache code called Unstructured Information Management Architecture, or UIMA, that was created by IBM back in 2005 to help put unstructured data into relational databases. The UIMA code is not just a framework for storing data, but it is also where the natural language processing (NLP) capabilities of the system reside. This is what allows Watson to take in textual questions and give back textual answers. Question analysis was done using Prolog, and the question parsing is one of the key things that makes Watson work (or not work, as was the case early on.) Some of Watson’s algorithms were done by hand in C and C++ for performance reasons. Building the original Watson system took four years and 27 researchers–who, IBM openly admits, stood on the shoulders of countless researchers working over decades in artificial intelligence, semantics, and related fields.
To commercialize the Watson system will require it to have other capabilities, including the ability to hear and to see (processing both pictures and video). IBM has been working on these functions for the past several years, and it is not clear what state of development these senses are in. The Watson system will also need to be able to parse datasets and draw pictures, and IBM has been working on this as well. The goal is to expand outwards from initial dabblings in medicine, where IBM has had several partnerships, into other fields such as finance, retail, and travel. No matter what field, the basic idea is the same: There is too much information scattered around for people to make informed decisions for their work or their life, and a system like Watson can offer up suggestions like a human expert would.
Expertise in specific industries is the key to success with such cognitive systems, and that is one of the reasons why IBM has chosen Michael Rhodin, who was formerly the senior vice president of IBM’s Software Solutions Group, to head up the new Watson business unit. The group will be located at 51 Astor Place, in a brand new building that is at the heart of Silicon Alley in the Big Apple, and it will have several hundred employees to start with, and over 2,000 are expected to be in the group by the end of the year. The Watson Group has been given a budget of more than $1 billion over the next few years. (IBM was not more specific about the term of the investment.) That budget includes $100 million in venture capital to invest in startups who want to leverage Watson technologies to create applications for specific industries.
Last November, IBM did a soft launch of a cloudy implementation of Watson, giving software developers access to Watson’s APIs so they could pump their data into the system and have it do question-answer processing. That Watson cloud has drawn over 890 companies and individual developers so far. IBM had already launched an application called Watson Engagement Advisor, which puts the system into call centers to help the humans in customer service answer questions better. It is not hard to imagine voice processing extensions that would allow such a program to replace the telephone messaging systems and a lot of the people in call centers, much as companies try to do with web-based tech support these days.
Three new Watson-based applications were announced today. They include Watson Analytics, which was developed under the codename “Neo” and which will go into beta next week. This analytics module uses natural language to help users find datasets and, based on input, can prep the data, find the most important relationships, and present it visually, according to Rhodin. Another new module in development is called Watson Explorer, which is a data discovery, search, and navigation service. Watson Discovery Advisor is the third new one that was developed in conjunction with publishers and pharmaceutical companies to sift through the large amount of material that gets published. (In the medical field alone, there are over 5,000 articles published each month, IBM says, and no one can keep up.)
As part of the announcement today, IBM is promising to port Watson over to its SoftLayer public cloud, and that either means moving Watson off Power Systems iron – which was chosen originally because of its high memory and I/O bandwidth compared to X86 alternatives – or putting Power machines inside the SoftLayer cloud. The original Watson system, by the way, had 90 quad-socket server nodes with 16 TB of main memory. IBM says it can now get the same capability into three servers and deliver a factor of 24X improvement in performance.