New A.I. application can write its own code
Computer scientists have created a deep-learning, software-coding application that can help human programmers navigate the growing multitude of often-undocumented application programming interfaces, or APIs.
Designing applications that can program computers is a long-sought grail of the branch of computer science called artificial intelligence (AI). The new application, called Bayou, came out of an initiative aimed at extracting knowledge from online source code repositories like GitHub. Users can try it out at askbayou.com.
“People have tried for 60 years to build systems that can write code, but the problem is that these methods aren’t that good with ambiguity,” says Bayou co-creator Swarat Chaudhuri, associate professor of computer science at Rice University. “You usually need to give a lot of details about what the target program does, and writing down these details can be as much work as just writing the code.”
“Bayou is a considerable improvement,” he says. “A developer can give Bayou a very small amount of information—just a few keywords or prompts, really—and Bayou will try to read the programmer’s mind and predict the program they want.”
Chaudhuri says Bayou trained itself by studying millions of lines of human-written Java code. “It’s basically studied everything on GitHub, and it draws on that to write its own code.”
Bayou co-creator Chris Jermaine, a professor of computer science who co-directs Rice’s Intelligent Software Systems Laboratory with Chaudhuri, says Bayou is particularly useful for synthesizing examples of code for specific software APIs.
“Programming today is very different than it was 30 or 40 years ago,” Jermaine says. “Computers today are in our pockets, on our wrists and in billions of home appliances, vehicles, and other devices. The days when a programmer could write code from scratch are long gone.”
Bayou architect Vijay Murali, a research scientist at the lab, says, “Modern software development is all about APls. These are system-specific rules, tools, definitions, and protocols that allow a piece of code to interact with a specific operating system, database, hardware platform, or another software system. There are hundreds of APIs, and navigating them is very difficult for developers. They spend lots of time at question-answer sites like Stack Overflow asking other developers for help.”
Murali says developers can now begin asking some of those questions at Bayou, which will give an immediate answer.
“That immediate feedback could solve the problem right away, and if it doesn’t, Bayou’s example code should lead to a more informed question for their human peers,” Murali says.
Jermaine says the team’s primary goal is to get developers to try to extend Bayou, which has been released under a permissive open-source license.
“The more information we have about what people want from a system like Bayou, the better we can make it,” he says. “We want as many people to use it as we can get.”
The researchers based Bayou on a method called neural sketch learning, which trains an artificial neural network to recognize high-level patterns in hundreds of thousands of Java programs. It does this by creating a “sketch” for each program it reads and then associating this sketch with the “intent” that lies behind the program.
When a user asks Bayou questions, the system makes a judgment call about what program it’s being asked to write. It then creates sketches for several of the most likely candidate programs the user might want.
“Based on that guess, a separate part of Bayou, a module that understands the low-level details of Java and can do automatic logical reasoning, is going to generate four or five different chunks of code,” Jermaine says. “It’s going to present those to the user like hits on a web search. ‘This one is most likely the correct answer, but here are three more that could be what you’re looking for.’”
The Defense Advanced Research Projects Agency funded the initiative that led to the project.
The researchers will present a paper on Bayou on May 1 in Vancouver, British Columbia, at the Sixth International Conference on Learning Representations.