What is Watson?
Watson is an IBM supercomputer named after Thomas Watson, the “father” of the modern IBM.
It uses AI (natural language processing and machine learning) to answer questions and shot to fame when it defeated human champions on the US quiz show Jeopardy in 2011.
- Jeopardy is a backwards quiz, where answers are provided and must be matched back to a question that would fit the answer.
- Questions are often cryptic and use puns and other word tricks, so it was a good test for a natural language computer.
In its initial quiz-answering form the system accessed 90 servers which held 200 million pages (only 4 Tb) of information (Watson doesn’t need the internet).
- It did however store all the information in RAM in order to be as fast as the human competitors.
- Watson took three seconds to respond on average, with a range of 1 to 6 seconds.
It also has a machine learning element, as right and wrong answers are used to inform future responses.
Ken Jennings, one of the human champions defeated by Watson, later wrote:
Just as factory jobs were eliminated in the 20th century by new assembly line robots, Brad and I were the first knowledge industry workers put out of work by the new generation of thinking machines. Quiz show contestant may be the first job made redundant by Watson, but I’m sure it won’t be the last.
Size and the cloud
At the time it won Jeopardy, Watson was still a big beast (the size of 10 refrigerators) and appeared to have a somewhat “brute force” approach to problem solving.
- In fact the system pursues many possible solutions in parallel and weights them with confidence ratings, choosing the answer it is most confident about.
It was based on 90 IBM 3.5 Ghz Power 750 parallel processing servers with 2,880 processor threads and 16 Tb of RAM.
- This configuration is not that powerful – Watson runs at 80 TeraFLOPs, well outside the top 1000 computers in the world.
- Nevertheless, Watson’s initial success was based more on rapid execution of many algorithms simultaneously (hardware) than on new algorithms (software).
As with all things computing, its physical size is shrinking all the time, and its processing power is increasing.
- At first, Watson was a physical installation at IBM headquarters, but the system has since been connected to the cloud via IBM’s BlueMix platform.
- Companies (hospitals, law firms) are now able to rent time on the machine.
IBM has also released a set of APIs (application programming interfaces) which now number 40 and allow firms to integrate Watson’s capabilities into their own systems.
- As well as natural language processing, these include sentiment analysis (looking at things like tweets to assess mood) and personality analysis via a person’s online output.
What’s so good about Watson?
The history of artificial intelligence is full of “breakthrough” moments that have failed to live up to their potential.
- Like trading systems optimised to fit past data, which don’t work going forward, AI systems are usually narrowly based, and not able to be adapted towards broader real-world problems.
Watson has two strengths in this regard:
- it processes unstructured data through a process known as text mining or cognitive computing
- and it can interact with humans through natural speech
Basically this means that Watson should be good at answering people’s questions, and giving them advice.
- It’s also likely that systems like Watson will eventually discover new insights that have not been found by human researchers.
Watson will be particularly useful in areas where vast amounts of unstructured data need to be processed.
- Humans are not good at handling vast amounts of information, and suffer from an anchoring bias (eg. with doctors, from an initial diagnosis).
Eighty percent of the world’s data is unstructured, and in theory Watson has applications in any field where there is a lot of unstructured text.
An obvious example is legal analysis, for which a derivative of Watson (ROSS) is used to analyse around a billion text documents.
- A help-desk system is also in place in a healthcare insurance provider.
Watson is also expected to be used by banks and insurance companies.
- Not only could it take over customer and internal help desks, it could also be used to detect credit card fraud and perhaps to act as a robo-advisor on investment portfolios.
- Implementing a system in the banking environment is hard to do at small scale, and no bank has yet been willing to pay up for a full-scale trial.
- Citibank signed an initial agreement in 2012, but nothing came of it.
Another interesting field could be weather forecasting, and particular weather-related business risks.
Watson is also working with chefs from the Institute of Culinary Education (ICE) to try to discover new flavour combinations.
- A cookbook with recipes “inspired by Chef Watson” is planned.
It can only be a matter of time before Watson-like capability is available as an assistant app on a smart phone.
- Shoppers would be able to access natural language assistance while in physical stores.
Critics say that IBM has used the Jeopardy win to surround a set of non-revolutionary technologies with a halo effect in order to shift product.
- There may be something in this, but the proof of the pudding is in the eating, and if Watson (or something called Watson) can be used to crack real-world problems – areas where there may be no “right answer” – it’s still a victory for AI.
Branding can open doors, and not every project has to be a moonshot (like “ending cancer”).
- Unlike Google and Facebook, who are basically using AI internally to improve their existing services, IBM is at least trying to open up AI to the wider business world.
- IBM is behind both Google and Facebook in terms of access to “generic” consumer data, but is ahead in domain specific knowledge.
Watson the doctor
Another area where Watson is being used is in medical diagnosis, specifically lung, prostate and breast cancer.
- Watson has access to 600,000 pieces of medical evidence, two million pages of medical journals and 1.5 million patient records.
- It has been estimated that it would take a doctor 160 hours per week to read all the new medical knowledge that is published.
- In clinical trials, Watson had a 90% successful diagnosis rate for lung cancer, compared to 50% for human doctors.
Not only is Watson better than people , it will be cheaper as well.
- Thirty per cent of the $2.3 trn spent on healthcare in the US is thought to be wasted, and Watson will be used to improve efficiency.
- And because of its iterative learning capabilities, Watson will get better at working out how to treat patients, and also at how to save money while doing so.
Problems with the output
There are two issues in using expert systems like Watson:
- Since it is trained by human experts, and uses a probabilistic approach (running several approaches to a solution and weighting the results), it’s possible that it will make judgments that are similar to those of humans, and include human biases.
- It’s also the case that Watson produces potentially brilliant results without showing it’s working.
- In business this may not be enough.
- Radical changes of direction and strategy – even simple efficiency moves like losing a factory – required detailed justification which Watson can’t at the moment provide.
IBM is aware of this and is working on a visualisation tool called WatsonPaths which will attempt to explain how Watson reached its conclusion.
What can’t it do?
Watson also has contextual issues.
You may remember the Microsoft chatbot release on Twitter in 2016.
- It rapidly adopted the offensive views of other Twitter users, and the plug had to be pulled after one day.
Watson is like that, too.
In 2015, the Urban Dictionary (of current – particularly internet – slang) was added to Watson’s database.
- Though Watson could understand the colourful definitions, it struggled to grasp the difference between polite and offensive speech.
- The experiment had to be abandoned.
Until next time.