Sunday, April 29, 2018

Questions Americans Want Answered: "Does my algorithm have a mental-health problem? "

From Aeon:
Is my car hallucinating? Is the algorithm that runs the police surveillance system in my city paranoid? Marvin the android in Douglas Adams’s Hitchhikers Guide to the Galaxy had a pain in all the diodes down his left-hand side. Is that how my toaster feels?

This all sounds ludicrous until we realise that our algorithms are increasingly being made in our own image. As we’ve learned more about our own brains, we’ve enlisted that knowledge to create algorithmic versions of ourselves. These algorithms control the speeds of driverless cars, identify targets for autonomous military drones, compute our susceptibility to commercial and political advertising, find our soulmates in online dating services, and evaluate our insurance and credit risks. Algorithms are becoming the near-sentient backdrop of our lives.

The most popular algorithms currently being put into the workforce are deep learning algorithms. These algorithms mirror the architecture of human brains by building complex representations of information. They learn to understand environments by experiencing them, identify what seems to matter, and figure out what predicts what. Being like our brains, these algorithms are increasingly at risk of mental-health problems.

Deep Blue, the algorithm that beat the world chess champion Garry Kasparov in 1997, did so through brute force, examining millions of positions a second, up to 20 moves in the future. Anyone could understand how it worked even if they couldn’t do it themselves. AlphaGo, the deep learning algorithm that beat Lee Sedol at the game of Go in 2016, is fundamentally different. Using deep neural networks, it created its own understanding of the game, considered to be the most complex of board games. AlphaGo learned by watching others and by playing itself. Computer scientists and Go players alike are befuddled by AlphaGo’s unorthodox play. Its strategy seems at first to be awkward. Only in retrospect do we understand what AlphaGo was thinking, and even then it’s not all that clear.
To give you a better understanding of what I mean by thinking, consider this. Programs such as Deep Blue can have a bug in their programming. They can crash from memory overload. They can enter a state of paralysis due to a neverending loop or simply spit out the wrong answer on a lookup table. But all of these problems are solvable by a programmer with access to the source code, the code in which the algorithm was written.

Algorithms such as AlphaGo are entirely different. Their problems are not apparent by looking at their source code. They are embedded in the way that they represent information. That representation is an ever-changing high-dimensional space, much like walking around in a dream. Solving problems there requires nothing less than a psychotherapist for algorithms.

Take the case of driverless cars. A driverless car that sees its first stop sign in the real world will have already seen millions of stop signs during training, when it built up its mental representation of what a stop sign is. Under various light conditions, in good weather and bad, with and without bullet holes, the stop signs it was exposed to contain a bewildering variety of information. Under most normal conditions, the driverless car will recognise a stop sign for what it is. But not all conditions are normal. Some recent demonstrations have shown that a few black stickers on a stop sign can fool the algorithm into thinking that the stop sign is a 60 mph sign. Subjected to something frighteningly similar to the high-contrast shade of a tree, the algorithm hallucinates.

How many different ways can the algorithm hallucinate? To find out, we would have to provide the algorithm with all possible combinations of input stimuli. This means that there are potentially infinite ways in which it can go wrong. Crackerjack programmers already know this, and take advantage of it by creating what are called adversarial examples. The AI research group LabSix at the Massachusetts Institute of Technology has shown that, by presenting images to Google’s image-classifying algorithm and using the data it sends back, they can identify the algorithm’s weak spots. They can then do things similar to fooling Google’s image-recognition software into believing that an X-rated image is just a couple of puppies playing in the grass....MUCH MORE