The Curious Case of The Artificial Intelligence Accused of Cheating
By Jon Pearson, Sales Director, Transformational Digital Journeys, Banking @ Feedzai
Remember Deep Blue?
This marvel of modern technology (back in 1996, at least) beat chess world champion Garry Kasparov in the first game of a six game match, although it lost the match, 4-2. It then went on to win a re-match (3½–2½) with Kasparov one year later.
Kasparov suggested that Big Blue had cheated. In particular, the grandmaster pointed to an uncharacteristic move by Deep Blue in the first game of the rematch, an incident that unsettled Kasparov and affected his performance afterwards.
The controversy never fully got resolved at the time. But IBM’s share price (“Big Blue”), the developer of Deep Blue, rose significantly. One estimate puts the increase at 20%.
Now, 20 years later, computers have graduated to playing Go, a game more complex than chess and a favorite in many cultures. Google DeepMind and its AlphaGo software beat world Go champion Lee Se-dol earlier this year.
Deep Blue for chess, AlphaGo for Go, and Watson (in 2011) for the TV game show Jeopardy! All are notable examples of machines successfully beating their expert human counterparts.
In 2014, the “Eugene Goostman” chatbot from Princeton University was claimed to have fooled 33% of its human correspondents into believing Eugene was human. The more human beings work towards the same goal, the closer they (and humanity) get to that goal. That’s the veritable numbers game aspect of AI.
Alan Turing, creator of the Turing Test for machine intelligence, would surely have wanted to chat with Eugene. In 1950, Turing suggested a way to test a machine’s ability to behave like a human being. If a human evaluator could not distinguish between machine and human interactions, both being done via a text-only medium (keyboard and screen, for instance), then the machine passed the test.
It’s amazing to see how close we’ve come.
So what of Kasparov, pitting his neurons against Deep Blue’s code, and accusing it of cheating?
Any use by IBM of human chess players to help Deep Blue during a game would have contravened the rules of the match. IBM denied any cheating, although it used provisions in the rules allowing developers to tweak Deep Blue between games in the same match.
After Kasparov thought the machine had made a human-aided move, he asked for the machine logs to be released. IBM refused to do so at first, only making them available some time later. The company also retired and dismantled Deep Blue after its victory.
Per a WIRED article on this event, here is what actually happened:
“Seeing the machine’s failings in game one, he says, IBM went back to the drawing board, reassigning relative values for different features of the game. Whereas the pre-tweaked machine might have thought the queen’s mobility was more important than a captured pawn, the adjusted Deep Blue might calculate that a good fat pawn, in a particular context, looked juicier than a free-ranging queen.”
It’s what we call feature engineering in the machine learning and artificial intelligence domain. Essentially it is the process of using the domain knowledge of the data to tease out the important elements/variables that power the machine learning model. In this case the features tweaked had to do with the attributes of certain chess pieces. The right features make all the difference.
At Feedzai, we know this is the most critical step in creating machine learning models, it’s why we hire the best data scientists and why we built a machine learning platform that let us iterate quickly on these models.
And finally, what of the increased stock price?
Due to challenges in the speed, power and effectiveness of the algorithms, this was euphoria in 1996, but is now a rational investment in 2016.
The AI wars have begun; big consumer technology companies like Google, Facebook and Amazon are rapidly unleashing this power into the world. At Feedzai, we’re doing our part by using AI to enable safe and frictionless commerce.
The idea that artificial intelligence could exasperate a chess grandmaster, improve the fortunes of a global company and herald in a new age of AI would have made Alan Turing a proud father.
Other Related Resources You May Like:
Latest posts by Ken Bui (see all)
- EU Fifth AML Directive: How Banks Can Prepare for Five Key Changes - November 18, 2019
- AI Best Practices to Improve Enterprise Risk Outcomes - October 22, 2019
- Why It’s So Hard for Challenger Banks to Fight Financial Crime - September 24, 2019
Subscribe to stay infomed