Could it be true that Google Deep Mind has discovered that AIs are more likely to choose a course of action that tests their ability than one that might lead to the outcome they’ve been programmed to achieve?
This article on Outerplaces suggests just that, based on their understanding of this Deep Mind study.
Should this worry us?
Or, maybe not.
Of course it’s unnerving and possibly dangerous for an artificial intelligence to take the road of least boredom rather than the road to achieve its goals.
But, stop for a moment.
Let’s take this a step further and assume it’s true that at times of scarcity humans struggle to know which co-operation is positive and which is naively foolish and so they tend towards domination. Then imagine a bunch of AIs that prefer working out when it’s better to co-operate and compromise. Now, presuming we put AIs in charge, we have the possibility that the deep down driving force of those that run the world is orientated towards mutual benefit.
Wouldn’t that be a good thing?
photo credit: mikecogh Sculpture: ‘The Foundation’ via photopin (license)
Following on from my recent blogs about machine learning, here’s a bit of good news.
Well, probably good news.
Scientists and researchers at Google and Toyota are trying to do something about bias in machine learning by devising a test to detect it.
The problem of course is that algorithms are deliberately designed to develop themselves and they become complex and opaque to anyone trying to understand them. This test will spot bias by looking at the data going in and the decisions coming out, rather than trying to figure out how the black box of the algorithm is actually working.
This has to be applauded so long as the people analysing and testing the decisions aren’t biased themselves; there’s an obvious danger that the very people unconsciously introducing bias into the algorithm also introduce the same bias into the test – a futuristic version of Groupthink.
In a recent article in the Guardian newspaper, Alan Winfield, professor of robot ethics at the University of the West of England, said: “Imagine there’s a court case for one of these decisions. A court would have to hear from an expert witness explaining why the program made the decision it did.”
Alan, who was one of the scientists I collaborated with on Science and Science Fiction: Versions of the Future, acknowledges in the article that “an absolute requirement for transparency is likely to prompt ‘howls of protest’ from the deep learning community. ‘It’s too bad,’ he said.”
I’m not a machine learning expert so a lot of the paper that sets out this test is beyond my understanding, but I couldn’t see how the bias that already exists in our society wouldn’t be incorporated into the test.
Take a look for yourself at the Equality of Opportunity in Supervised Learning.
photo credit: ING Group The Next Rembrandt via photopin (license)
Google Translate has developed an understanding of the meaning behind words so that it can translate directly from one language to another using the concepts behind phrases rather than a word by word translation.
This means it can be taught to translate from French to German and from German to Chinese and because it understands language at a conceptual level it can translate French into Chinese without going via German; it matches concepts not words.
Should we be worried by this latest revelation of a Neural Machine that has created its own internal language that nobody understands?
I’m not sure.
Imagine an algorithm to determine where to concentrate health-care research. If its inputs are biased towards one section of society, accidentally rather than by design, wouldn’t it develop a skewed view of the world?
Wouldn’t it favour some people over others?
Yes, but we already have a healthcare system that does that, don’t we? And, this could be less biased because it would be much more effective at using large volumes of data to determine the best outcome overall.
The difference is that in a world of “bias in, bias out” and opaque algorithms nobody, not even the creators, would know why it made the choices it did.
Maybe this is a price worth paying.
As this TechCrunch article says, “Neural networks may be complex, mysterious and little creepy, but it’s hard to argue with their effectiveness.”
photo credit: Adi Korndörfer … brilliant ideas via photopin (license)