Programming by Examples (PBE) involves synthesizing intended programs in an underlying domain-specific programming language (DSL) from example-based specifications. This new frontier in AI enables computer users, 99% of whom are non-programmers, to create scripts to automate repetitive tasks. PBE can provide 10-100x productivity increase for data scientists, business users, and developers for various task domains like string/number/date transformations, structured table extraction from log files/web pages/PDF/semi-structured spreadsheets, transforming JSON from one format to another, repetitive text editing, repetitive code refactoring and formatting. PBE capabilities can be surfaced using GUI-based tools, code editors, or notebooks, and the code can be synthesized in various target languages like Java or even PySpark to facilitate efficient execution on big data.

There are three key components in a PBE system. (i) A search algorithm that can efficiently search for programs that are consistent with the examples provided by the user. We leverage a divide-and-conquer based deductive search paradigm that inductively reduces the problem of synthesizing a program expression of a certain kind that satisfies a given specification into sub-problems that refer to subexpressions or sub-specifications. (ii) Program ranking techniques to pick an intended program from among the many that satisfy the examples provided by the user. (iii) User interaction models to facilitate usability and debuggability.

Each of these PBE components leverage both symbolic reasoning and heuristics. We make the case for synthesizing these heuristics from training data using appropriate machine learning methods. In particular, we use neural-guided heuristics to resolve any resulting non-determinism in the search process. Similarly, our ML-based ranking techniques, which leverage features of program structure and program outputs, are often able to select an intended program from among the many that satisfy the examples. Finally, Our active-learning-based user interaction models, which leverage clustering of input data and semantic differences between multiple synthesized programs, facilitate a bot-like conversation with the user to aid usability and debuggability. That is our algorithms that deeply integrate neural techniques with symbolic computation can not only lead to better heuristics, but can also enable easier development, maintenance, and even personalization of a PBE system.