TDD-ML: Test-Driven Development for ML Systems

Abstract

What we aim to do in this project is develop a methodology and a tool for analysis of test data sets used in ML model development and acquisition processes. We will then integrate into a MLOps pipeline to support maintenance and expansion of the test sets in a production environment. Specifically, we will develop a methodology and tool were:1. Representative ML model developers, when using the tool in conjunction with software engineering TDD practices, will better specify the ML model, resulting in an 30%increase in model accuracy, in line with results of Afan, et al. 2022.2. Use of the tool during ML model evaluation and acquisition will lead to the selection of an ML model that will perform at least 20% better in production, in line with the findings of Oakden-Rayner, et al. 2020.3. Use of the tool, when integrated into a DoD MLOps pipeline, during ML model evaluation and acquisition will lead to the selection of an ML model that will not experience as large of a performance drop off when in production when compared to baseline.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Jan 06, 2022
Accession Number
AD1189826

Entities

People

  • Rachel Brower-sinning

Organizations

  • Carnegie Mellon University

Tags

Communities of Interest

  • Human Systems

DTIC Thesaurus Topics

  • Acquisition
  • Copyrights
  • Data Sets
  • Department Of Defense
  • Engineering
  • Environment
  • Guarantees
  • Materials
  • Pipelines
  • Production
  • Software Development
  • Statistical Tests
  • Students
  • Test And Evaluation
  • Test Sets
  • Universities
  • Validation

Fields of Study

  • Engineering

Readers

  • Computational Modeling and Simulation
  • Life Cycle Cost Analysis
  • Neural Network Machine Learning.