The promise of Deep Learning for artificial intelligence...
Artificial Intelligence (A.I.) is considered one of the biggest technological disappointments of the 20th century. During the early 1960s, many researchers thought we were just a decade or two away from building thinking machines rivaling that of the human brain, capable of writing music, analyzing literature, talking philosophy — you name it. And Hollywood bought into the concept hook, line and sinker, resulting in a spate of films pitting man against A.I.-powered machines, from the duplicitous HAL 9000 computer in 2001: A Space Odyssey to the murderous Skynet defense system in the Terminator films to the all-controlling software and hardware of The Matrix.
But not only did these scientific predictions not take shape, they were spectacularly wrong. By the late 1980s, the field had entered what’s known as “A.I. winter,” as funding was drying up, new ideas weren’t really panning out and little progress was taking place. Hype followed by failure and criticism were generally the norm in the world of machine intelligence. Today, we understand that A.I.’s pioneers underestimated the complexity of the problem by many orders of magnitude. Computers are great at a lot of things, but they can’t grasp the basic tasks routinely performed by a 2-year-old child, a cat or even a honey bee.
Things may be starting to change, however. A new type of machine learning called “deep learning” is showing a great deal of promise. This new computer algorithm far exceeds its predecessors in its ability to recognize symbols and images, making self-driving cars and robotic butlers a real possibility in the coming decades. Useful digital assistants, like Apple’s Siri, are already using deep learning techniques. They are still limited, but what they can do was unthinkable just a few years ago, and researchers say the technology is now advancing at an unprecedented pace.
Deep learning is based on neural networks, simplified models of the way clusters of neurons act within the brain that were first proposed in the 1950s. The difference now is that new programming techniques combined with the incredible computing power we have today are allowing these neural networks to learn on their own, just as humans do. The computer is given a huge pile of data and asked to sort the information into categories on its own, with no specific instruction. This is in contrast to previous systems that had to be programmed by hand. By learning incrementally, the machine can grasp the low-level stuff before the high-level stuff. For example, sorting through 10,000 handwritten letters and grouping them into like categories, the machine can then move on to entire words, sentences, signage, etc. This is called “unsupervised learning,” and deep learning systems are very good at it.
Impressive, but we still have a long way to go. This is only a small step in building a truly intelligent machine. Deep learning has no way of establishing causal relationships or making logical inferences. IBM’s amazing, Jeopardy-winning Watson system uses deep learning only as part of a much larger ensemble of techniques, including statistical analysis, Bayesian inference and deductive reasoning. We still aren’t even close to creating HAL 9000 or anything truly intelligent. However, if the pace of progress continues, self-driving cars may become commonplace and machines just might start to interact with humans in a way that doesn’t result in one’s iPhone being thrown against the wall.