Google Brain researchers have created an automated coding tool for machine learning frameworks like TensorFlow, that outperforms humans on challenging development tasks.
Google's TF-Coder synthesizes tensor manipulation programs from input and output examples and natural language descriptions, using per-operation weights to enumerate over TensorFlow expressions in order of increased complexity; a novel type- and value-based filtering system oversees constraints imposed by the TensorFlow library.
TF-Coder weighs 134 tensor-manipulation operations of the 500 in TensorFlow, and accommodates problems involving compositions of four or five different operations and data structures of 10 or more components.
The researchers said TF-Coder realized "superhuman" performance on a range of real problems from question-and-answer website StackOverflow, completing tasks that human programmers would spend minutes to hours on.
The team said, "TF-Coder can help both machine learning beginners and experienced practitioners in writing tricky tensor transformation programs that are common in deep learning pipelines."
View Full Article
Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA