While expert systems rose and fell, another approach was quietly gaining strength. Instead of writing rules by hand, researchers increasingly built systems that learned patterns from data. By the 2000s, machine learning was powering search engines, spam filters and recommendations, often without being called AI at all.
Backpropagation
Minsky and Papert had shown the limits of single-layer perceptrons. Multi-layer networks could overcome them, but how could they be trained? The answer was an algorithm called backpropagation, which calculates how much each connection in a network contributed to an error and adjusts it accordingly, working backward from the output. Versions had been described earlier, but in 1986 a paper by David Rumelhart, Geoffrey Hinton and Ronald Williams in the journal Nature showed clearly how it allowed multi-layer networks to learn useful internal representations. Interest in neural networks revived, a movement then called connectionism.
Reading handwriting
In 1989, Yann LeCun and colleagues at Bell Labs applied backpropagation to a special kind of network, the convolutional neural network, inspired by the visual cortex, which scans images for local patterns such as edges and curves. Their system learned to read handwritten digits on postal zip codes. A later version was used commercially to read handwritten amounts on bank checks, processing a large share of checks in the United States by the late 1990s. It was one of the first major real-world successes of neural networks.
Statistics wins
Through the 1990s, many researchers turned to statistical methods. In speech recognition, systems at IBM and elsewhere used probability models trained on recorded speech rather than linguists' rules. A remark attributed to the IBM researcher Fred Jelinek captures the mood: that every time he fired a linguist, the performance of his speech recognizer went up. In 1988, the computer scientist Judea Pearl published work on Bayesian networks, which represented uncertain knowledge with probabilities, transforming how AI handled uncertainty. Techniques such as decision trees and support vector machines became standard tools. AI shifted from trying to tell computers how to think toward letting them find patterns in examples.
Deep Blue
The most famous AI event of the 1990s, however, used a different method. In May 1997, IBM's chess computer Deep Blue defeated the world champion Garry Kasparov in a six-game match, three and a half points to two and a half. It was the first time a computer had beaten a reigning world champion in a match under standard conditions. Deep Blue relied mainly on brute-force search, evaluating around two hundred million positions per second with specialized hardware, combined with chess knowledge programmed by experts. Kasparov suspected human interference at one point, which IBM denied. The victory was a milestone, but many researchers noted that Deep Blue could do nothing but play chess.
Learning to play by itself
In 1992, Gerald Tesauro at IBM created TD-Gammon, a neural network that learned to play backgammon at the level of top human players mainly by playing against itself, using a method called reinforcement learning, in which a system learns from rewards and penalties. It foreshadowed later breakthroughs in games.
AI everywhere, quietly
By the 2000s, machine learning was part of everyday life. Spam filters learned to recognize junk email from examples. Online stores and streaming services used recommendation systems. From 2006 to 2009, the company Netflix offered a one million dollar prize to anyone who could improve its movie recommendation algorithm by ten percent; a team of researchers won in 2009 by combining many models. Credit card companies used machine learning to detect fraud. Because these systems worked, people stopped calling them AI, an effect sometimes summed up in the quip that AI is whatever has not been done yet.
Self-driving cars
In 2004, the US defense research agency DARPA held a Grand Challenge: a race for self-driving vehicles across the Mojave Desert. No vehicle finished; the best traveled about seven miles. In 2005, five vehicles completed a 132-mile course, and Stanford's car, Stanley, which used machine learning to interpret its sensors, won. The challenges launched the modern self-driving car industry.
Watson
In February 2011, IBM's Watson system defeated two of the greatest champions of the television quiz show Jeopardy!, Ken Jennings and Brad Rutter, answering questions posed in natural language with wordplay and puns. Watson combined many techniques, including statistical language processing. IBM later tried to apply Watson to medicine, with disappointing results that showed how hard it is to move from games to real-world problems.
Worked example: rules versus learning
Compare two ways to build a spam filter.
- Rules. An engineer writes rules: block emails containing “free money” or many exclamation points. Spammers quickly change their wording. The engineer must keep writing new rules, and the filter blocks legitimate emails that happen to match.
- Learning. The system studies thousands of emails labeled spam or not spam and learns which features predict spam. When spammers change tactics and users mark new spam, the system learns again.
- Trade-offs. Learning adapts better but needs lots of labeled data, and it can be hard to explain why it made a particular decision. It can also learn the wrong lessons if its data is biased.
This shift, from hand-written rules to learned patterns, is the most important turn in AI's history. It made modern AI possible, and it created the challenges of data, transparency and bias that define AI debates today.