Posted November 6, 2019
Extracting Types from Python Machine Learning Libraries
Team Members and Titles

Ana Milanova (Associate professor, RPI)

Ingkarat (Bill) Rak Amnouykit (PhD student, RPI)

Andrew Ma (Co-terminal Masters student, RPI)

Daniel McCraven (Undergraduate researcher, RPI)

Martin Hirzel (Research staff member and manager, IBM Research)

Julian Dolby (Research staff member, IBM Research)

Summary

Our work aims to develop novel AI Automation techniques and tools that (1) improve productivity of AI/ML researchers and practitioners and (2) improve correctness, robustness, and maintainability of AI/ML software. Demand for AI/ML software development exceeds supply, and therefore, building software tools is increasingly important and relevant.

Key Findings

Thrust 1 will build type inference tools for Python, including standard types (int, float, string, list, etc.) and types characteristic of AI/ML (categoricals, tensors).

Thrust 2 will build precondition inference tools, based on Thrust 1, which can extract specifications for conditional hyperparameters of machine-learning operators.

Thrust 3 will use the results of Thrust 1 and Thrust 2 for error reporting in IDEs and for search space pruning in AI Automation tools.

Project Start Date
Project End Date