Federal lawmakers want to have a say in defining artificial intelligence. Researchers are now using TV shows to feed the predictive capability of an AI system. Google said in recent days it’s opening an AI-focused research facility in China. And on and on the headlines keep coming, all of which is to say that interest in AI remains acute — and its presence pervasive — as 2017 draws to a close.
And, based on a few recent developments, 2018 should be another big year of AI-related leaps forward as machines expand their influence over the minutiae of our lives.
Dr. Lea Frermann certainly thinks that’s the case. She’s a postdoctoral researcher at the University of Edinburgh who was part of a team that used episodes from the popular TV crime series CSI — specifically, the video, audio and text related to those episodes — to train an AI system to learn how to identify the bad guy correctly in each episode.
The researchers, Frermann explained to BGR, wanted to see if — given enough data — an AI system can be trained to solve problems that are difficult for humans. The result? The computer learned to pick out the likely perpetrator in each episode by the end 60 percent of the time, compared to an 85 percent success rate of human participants watching the show.
Frermann said those results are more positive than not for researchers like her. They were published at the end of October in the academic journal “Transactions of the Association for Computational Linguistics,” and the team hopes that it invites the rest of the academic community to build on what they found.
“In this paper,” the team writes, “we argue that crime drama exemplified in television programs such as CSI:Crime Scene Investigation is an ideal testbed for approximating real-world natural language understanding and the complex inferences associated with it.” The team, according to Frermann, was also largely pleased with the results they were able to get by training the AI system via the CSI episodes.
We’ll resist the urge to speculate on what Elon Musk thinks about this, about possibly hastening the arrival of an evil super-intelligence by training it with scripted crime drama. On a more serious note, the experiment — which generated plenty of headlines for the novel nature of the exercise and the TV show involved — also revealed something to the team that’s not as sexy but that shouldn’t be overlooked as we head into a new year that will bring us scores of new headlines about AI and the future.
You can’t get to perfect or even anywhere close to near-perfect accuracy in a system like the one Frermann’s team was using just by feeding it a copious amount of data.
“I don’t think only more data would help solve the whole issue” of making the rate of successful predictions even better, Frermann said. “I think we’d need to think really hard about what other information to give the model access to. Humans have so much world knowledge that’s hard to encode in a machine in an exhaustive way. I think we’d need to equip models with much better background knowledge and much better inference strategy. I don’t think just throwing unstructured data, more and more of it, would get us a long way farther. It might help a little bit.”
It’s worth watching what researchers do moving forward with insights like this. Which is one reason why 2018 may be a pretty interesting one in terms of AI news.
The federal government may also ensure that’s the case as well.
U.S. Sen. Maria Cantwell, a Democratic senator from Washington state, has drafted legislation that would set up a committee focused on artificial intelligence. The committee would be put in place by the Commerce Secretary and would advise that cabinet member about all things AI.
The language of the bill comes right out and says it: “It is the sense of Congress that understanding and preparing for the ongoing development of artificial intelligence is critical to the economic prosperity and social stability of the United States; as artificial intelligence evolves, it can greatly benefit society by powering the information economy, fostering better informed decisions and helping to unlock answers to questions that, as of the date of the enactment of this Act, are unanswerable.”
So there we have it. Congress wants to define AI and help the government harness and benefit from it. We have researchers using TV shows to see how well they can train an AI system. It will be interesting, to say the least, to see what 2018 holds.