Human vs. computer, and you can’t spot the difference
Trio including U of T rep teaching machines to think just like their creators
Pretend you’ve never seen a pineapple. Now imagine somebody shows you a pineapple.
After seeing just one, it would be very easy for you to identify a pineapple in a bowl of other fruits, draw something that kind of looks like a pineapple, tell the difference between its leaves and its body, or sketch something inspired by a pineapple but slightly different.
These tasks are so easy for humans that we don’t notice performing them: generalizing from a single example is a basic fact of cognition. But despite incredible leaps in artificial intelligence in recent years, machines have not been able to achieve such “one-shot” learning.
A trio of scientists from New York University, the University of Toronto and MIT have created a computer model that not only succeeds on these tasks, but repeatedly passed a “visual Turing test” — in other words, performing in a way that is indistinguishable from human action.
“We want to better understand how people learn — to us that means reverse-engineering how learning works in a human mind — and we also want to engineer or build machines that learn in more humanlike ways,” says MIT’s Joshua Tenenbaum, who co-authored the paper with Brenden Lake at NYU, and Ruslan Salakhutdinov at U of T. “We believe we’ve made an important step here.”
The researchers challenged the computer to perform a simple task: identify, parse and copy handwritten characters from alphabets around the world. After seeing a single Tibetan letter, for example, the algorithm could pick out other examples of that character drawn in different hand- writing, identify the strokes that make up the letter and redraw it, and generate made-up letters similar to a set of these characters.
Matching human performance at drawing Tibetan squiggles may not declare to the masses that futurists’ so-called singularity — genuine artificial intelligence — is nigh.
But the scientists’ approach, known as Bayesian program learning and described in a paper published in the journal Science, represents a remarkable advance in the drive to mimic aspects of human cognition with computer systems — one with far-reaching applications. The best machine-learning algorithms require hundreds or thousands of examples to operate successfully.
“Every week it seems we read about machines that can perform tasks in object recognition, face recognition or speech recognition seemingly as well as humans do,” said Tenenbaum. “Yet to scientists like me who study the mind, the gap between machine learning and human learning capacities remains vast. We want to close that gap, and that’s our long-term goal.”
Geoffrey Hinton, an artificial intelligence pioneer who works at U of T and Google, called the research “very impressive,” and said that the model’s ability to pass a visual Turing test is significant. “It’s quite an achievement to make that work.”
Hinton is the forefather of “deep learning,” another machine-learning approach that has achieved significant success and widespread adoption in recent years. To identify an object or translate human speech, deep learning improves as the number of examples it has seen goes up. The deep-learning algorithms that support applications like Google’s image search or Facebook’s face rec- ognition have likely seen millions of examples in order to “learn.”
The paper says that Bayesian program learning outperforms deep learning. But its authors and Hinton say that the two approaches succeed at different types of tasks, and both can improve by borrowing from the other — and that greater successes could come from creating hybrid systems. Deep learning shines where there is lots of data, even if that data is messy, whereas Bayesian program learning has shown it succeeds with limited, but very clean, data.
Yet Hinton said one of the most exciting outcomes of the new approach is its potential to silence critics who say that the way intelligent computer systems learn is nothing like how humans learn. The inability of computers to generalize from a single example was one of the mainstays of that debate, Hinton said.
“It’s the death of one more argument about why computer models aren’t good models of humans.”
The research is a remarkable advance in getting computers to mimic aspects of human cognition