Robot Psychologist Scientists Go Inside Minds of Machines.'
Career of the Future: Robot Psychologist
Engineers are using cognitive psychology to figure out how AIs think and make them more accountable
By Christopher Mims July 9, 2017 9:00 a.m. ET
Artificial-intelligence engineers have a problem: They often don’t know what their creations are thinking.
As artificial intelligence grows in complexity and prevalence, it also grows more powerful. AI already has factored into decisions about who goes to jail and who receives a loan. There are suggestions AI should determine who gets the best chance to live when a self-driving car faces an unavoidable crash.
Defining AI is slippery and growing more so, as startups slather the buzzword over whatever they are doing. It is generally accepted as any attempt to ape human intelligence and abilities.
One subset that has taken off is neural networks, systems that “learn” as humans do through training, turning experience into networks of simulated neurons. The result isn’t code, but an unreadable, tangled mass of millions—in some cases billions—of artificial neurons, which explains why those who create modern AIs can be befuddled as to how they solve tasks.
Most researchers agree the challenge of understanding AI is pressing. If we don’t know how an artificial mind works, how can we ascertain its biases or predict its mistakes?
As artificial intelligence gets more complex and prevalent, it also becomes more powerful. There are suggestions AI should determine which driver gets the best chance to live in the event of a crash of self-driving cars. It is important to know when AI will behave unexpectedly.
We won’t know in advance if an AI is racist, or what unexpected thought patterns it might have that would make it crash an autonomous vehicle. We might not know about an AI’s biases until long after it has made countless decisions. It’s important to know when an AI will fail or behave unexpectedly—when it might tell us, “I’m sorry, Dave. I’m afraid I
“A big problem is people treat AI or machine learning as being very neutral,” said Tracy Chou, a software engineer who worked with machine learning at Pinterest Inc. “And a lot of that is people not understanding that it’s humans who design these models and humans who choose the data they are trained on.”
An example can be found on Google Translate. Ask it to translate “doctor” to Portuguese, and it always returns the male form of the noun, médico, over the female, médica. Type in “nurse” and you get enfermeira (female)—never enfermeiro (male).
Conspiracy? No, it is a natural consequence of biases inherent in the bodies of literature used to train translation systems. Something similar happens in data when researchers eliminate the category of race: Other data, such as where a person lives, correlate so strongly with race that they become unintentional proxies for it.
Unlike with humans, we can’t just ask a robot why it does what it does. Artificial intelligences can excel at narrow tasks, but even those that talk have introspective powers about on par with a cockroach.
It is a difficult enough problem to crack that the Defense Advanced Research Projects Agency, better known as Darpa, is funding researchers working on “explainable artificial intelligence.”
Here’s why we’re in this pickle: A good way to solve problems in computer science is for engineers to code a neural network—essentially a primitive brain—and train it by feeding it enormous piles of data. Once the AI has had enough time to chew through a bunch of images labeled “cat,” for example, it can reliably pick out pictures of a cat.
The tricky bit is that neural networks learn by altering their own innards. This is basically how your brain works, too. And like the connections between the 86 billion or so neurons in your brain, the precise way an AI “thinks” is incomprehensible.
Engineers call this the “interpretability” problem (as in, the lack of it) and refer to neural networks as “black boxes”—things we can stimulate and observe but whose insides we can’t understand.
Researchers at DeepMind Technologies Ltd., a subsidiary of Alphabet Inc., announced a novel way to get inside the minds of machines: treat them like human children.
To say engineers are using the techniques of cognitive psychology on AI isn’t an analogy. The team at DeepMind used exactly the same tests and materials psychologists use on children to tease out how their AI thinks, says David Barrett, a DeepMind research scientist who worked on the project.
Decades of research on unpacking the human brain through cognitive science may now be applied to machines, potentially unlocking a whole new avenue for understanding AI and making it accountable, he said.
A result of DeepMind’s research: We now know at least one of its AIs—a “one-shot learning model” designed to learn words after being exposed to them only once—is, surprisingly, solving problems the same way humans do. Like humans, it is identifying objects by shape, even though it wasn’t taught to, and even though there are other ways to identify random objects, such as color, texture or movement. Previously, how it learned was opaque.
Understanding is just the beginning of how we interact with artificial intelligences. The other half of robot psychology is what might be described as therapy—that is, changing an AI’s mind.
Because engineers typically create many versions of an AI when trying to discover the best one, the use of cognitive psychology could give engineers more power to choose the ones that “think” the way we want them to, Mr. Barrett said. Alternatively, we might find it’s better when AIs don’t think like us: We might learn something new about how to solve problems.
The upshot is that when we replace human decision-makers with artificial intelligences, AIs have the potential to be better, with fewer mistakes and more accountability, because their output is measurable and we might be able to trace exactly how they make decisions.
We ask humans to do this all the time—in a court of law, when dissecting a business decision—but humans are notoriously unreliable narrators. With machines, at last, we could have decision-makers whose every bias and fleeting impulse can be inspected and potentially altered.
Appeared in the July 10, 2017, print edition as 'Career of the Future: Robot Psychologist Scientists Go Inside Minds of Machines.'