Watson in your pocket: Supercomputer gets own apps



Continue reading page |1|2


If you could quiz Watson, IBM's all-knowing supercomputer, from an app on your phone, what would you ask it?


That is the question facing app developers now that IBM has shrunk its cognitive computer from the bedroom-sized monster that won the TV quiz show Jeopardy! in 2011 to the size of just three stacked pizza boxes. Mini Watsons can now easily be installed in data centres worldwide and made available as a cloud service to cellphone users.


Until now it has been unclear what type of apps would make best use of Watson's capabilities. It is no ordinary computer, answering complex questions using data mining and machine learning. On 28 April, IBM unveiled the 25 best app ideas in response to the challenge it issued at February's Mobile World Congress in Barcelona, Spain. Three of the ideas will be developed, making them among the first apps powered by Watson-in-the-cloud. IBM has set aside $100 million for Watson apps and has already invested in Fluid of San Francisco. Fluid is writing an app for outdoor clothing firm The North Face to advise hikers on the right gear for the conditions, based on product data, user reviews and expert websites.


Watson's ability to learn is what sets it apart from conventional supercomputers that only carry out superfast number crunching. Instead, Watson's role is to learn all it can about different subjects using data gleaned from a wide range of sources, including databases, encyclopedias, news stories, reviews, dictionaries, peer-reviewed research papers and textbooks. Watson's wealth of knowledge means that it will be able to answer questions beyond the scope of mere web searches.


For example, one of the shortlisted apps could be great news for any worried new parents, if it makes the final three. Developer Biovideo of San Antonio, Texas, wants to train Watson on neonatal and infant medical data from sources such as medical journals, the American Academy of Pediatrics and the UK National Health Service. "So a mother with a sick child at 4 am will be able to use Watson to ask what is wrong with her baby and get a 100 per cent accurate response using data from these trusted sources," the firm claims in its proposal.


Crucially, Watson's source data can be "unstructured" and so does not need converting and organising into narrow database-field categories. When a question is posed in natural language – currently English though more languages are planned – Watson's linguistic processor examines it in 120 different ways to work out what is being asked.


Reasoning algorithms begin to find hypothetical answers in the data, scoring them with ever greater confidence levels and ultimately returning its best possible answer. Watson has already put that capability to good use at a number of cancer hospitals in the US (see "Dr Watson will see you now").


But it is changes to Watson's hardware that make it ready to hit our smartphones. To play Jeopardy!, in which contestants are given an answer and have to work out the question, Watson was trained on 200 million pages of data. That machine comprised 10 server racks containing 2880 processor cores and 15 terabytes of RAM. Now, says Rob High, chief technology officer of the IBM Watson Group in New York, the latest version of Watson performs even better than the original with just 32 processor cores and 256 gigabytes of RAM.


This is made possible because Watson's processors run in parallel, while its operating system runs concurrent software routines, known as threads, within each of the parallel-processing cores, which lets it learn much faster and more efficiently than before.


The new mini Watson, being no bigger than three pizza boxes, can easily be slotted into racks in a cloud data centre. The idea is that every developer will give Watson a mountain of data from their chosen area for the computer to learn from. Once trained, the apps let users ask Watson questions from their phone, says High. Watson can adapt to demand too. "We can expand the Watson resource elastically depending on the number of people asking it questions," he says. This is done by making more processors and memory available at the data centre if an app proves surprisingly popular.


One app that might do well is an ultimate guide to New York City. Proposed by Ontodia, a developer there, it aims to vastly improve on the disconnected and unimpressive hits Siri and Google offer up. Training Watson on municipal, state, federal, tourist and commercial databases will provide answers to complex, natural language searches like "what is the average income of this apartment block, correlated with property taxes and construction activity over the last ten years?". The firm's aim is to give New Yorkers a detailed profile of any neighbourhood.


Azoft of Novosibirsk, Russia, meanwhile, wants to provide phone users with an animated, intelligent avatar they can ask for advice – with Sigmund Freud and Albert Einstein cited as examples. Watson, Azoft says, would be trained with all Freud's and Einstein's books, articles, speeches, letters and interviews and would answer questions just as they might – perhaps with accurate speech synthesis, too.



Continue reading page |1|2


If you would like to reuse any content from New Scientist, either in print or online, please contact the syndication department first for permission. New Scientist does not own rights to photos, but there are a variety of licensing options available for use of articles and graphics we own the copyright to.