Pushmeet kohli microsoft




















Although programming languages have become much more user-friendly over the years, learning how to program is still a major endeavor that most computer users have not undertaken. October 19—21, View over recorded sessions from the Microsoft Research Summit, where researchers and engineers across Microsoft, and our colleagues in academia, industry, and government will come together to discuss cutting-edge work that is pushing the limits of science and technology.

For example, imagine that a user has a list of names that she wishes to format in a specific way, as shown below. She provides a few input-output examples, and the system auto-fills the remaining outputs shown in light gray.

In cases where users have hundreds or thousands of input strings, this can be a major time saver. The system performs this task by generating a program in a domain specific language DSL. The user does not have to understand the details of the DSL, and in fact she generally does not see the program at all.

In our DSL, the correct program corresponding to the above example is:. There are two key challenges in program synthesis. First, there are trillions of possible programs in our expressive DSL, and the correct program has likely never been seen before by the system.

Previous approaches to solving this problem — most notably the FlashFill system in Microsoft Excel — have relied on hand-crafted rules and heuristics to guide the search for the desired program.

Our system, called RobustFill, leverages recent advances in deep learning to take a data-driven approach to program synthesis without the need for any hand-crafted rules. A sketch of our network is shown below, and the details are provided. Overall, our system achieved 92 percent accuracy on a set of real-world benchmarks. The ability to successfully train neural architectures for learning to program in a rich functional language FlashFill DSL not only marks an exciting breakthrough in neural program synthesis, but also is a small but noteworthy step towards achieving more general artificial intelligence.

John has a PhD in chemistry from the University of Chicago. Before that, he worked on molecular dynamics simulations of proteins and supercooled liquids. John leads the development of next-generation AlphaFold models. Kathryn completed her PhD in systems approaches to biomedical science at the University of Oxford.

She then worked on vehicle routing and geospatial data at Ocado Technology. This ranges from setting up data processing pipelines to enabling the use of externally developed scientific software. About Science. Discovery and understanding Like the astronomers who built and used powerful telescopes to expand our understanding of the universe, our Science team builds innovative AI and machine learning systems.

Team profile.



0コメント

  • 1000 / 1000