Earlier this year, at a special tournament in a casino in Pittsburg, four of the world’s top poker players took on a computer called Libratus. The artificial intelligence system was developed by Tuomas Sandholm, professor of computer science at Carnegie Mellon University, along with his PhD student Noam Brown.
Computers had been extremely successful at chess and Go in the past, but they are actually relatively easy to model because the computer has all the information: there are no secrets on a chess board. Poker is different: each player has two secret cards (they were playing Texas Hold’em), so other players have to infer what those cards might be from the bidding.
For the first few days, Libratus was terrible. Then suddenly it started winning, and winning big. It had learnt to bluff. The interesting thing is that Brown confessed “When I see the bot bluff the humans, I’m like ‘I didn’t tell it to do that. I had no idea it was even capable of doing that.’”
Put another way, this is real artificial intelligence: it learnt what was going on and taught itself new tactics. That ability is behind the new best name for the technology: machine learning.
What has all this got to do with us? IBC each year offers a prize for the best paper in its technical conference. This year it was won by four researchers from TV Globo in Brazil, for a great piece of work with the snappy title Big data for data journalism, enhanced business analytics and video recommendation at Globo.
Allied to the big data in their work was machine learning to process this data, to draw unexpected conclusions from it. The data journalism in the title of the paper is a project which ingests published data from government and public sources in Brazil, then reads through what it calls the data reservoir to identify potentially interesting news stories. “The goal of the data journalism process is to reveal a story that is relevant and was unknown before data mining was carried out,” according to the paper.
At NAB I talked to Richard Friedel, executive VP of Fox. He told me “I’m interested in machine learning – I don’t call it artificial intelligence. I could have it train a system and develop the best workflows. We could even possibly automate sports highlights in the future – find the action and put together the highlights packages.”
Other sports specialists have talked about using machine learning to identify the goals and key action points, lining them up for a craft editor to smooth. It is an intriguing possibility.
ITV in the UK has recently implemented a system to automate the production of trailers. Editors develop the creative concept for a campaign and the production of all the various versions happens automatically. Using this system, the company produces more than 1000 assets a month. At present this is based on rules and information entered into a form by the commissioner, but the options for machine learning are obvious.
I think we all agree that the best way to run an efficient business is to get machines to do the dull and repetitive tasks (for which they are ideally suited) and have people do the creative, investigative, problem-solving stuff. But if a machine can teach itself to play poker and beat the best, what else can it learn to do?
How do we maintain control of such a system? Remember that the programmer behind Libratus didn’t know how it had learnt to bluff. How do we know that the plans and workflows it develops actually are the best possible solution, if we cannot see the workings? And who owns the intellectual property that the machine learning system creates?
There seem to be real possibilities opened up by machine learning, and forward-looking broadcasters like TV Globo are putting toes into the water. It will be fascinating to see how machine learning transforms the management – and maybe even the creativity – of the media business over the coming years.
 |
Guest blog by:
|
