Resource Efficient Machine Learning (Budgeted ML)

Machine learning applications often requires achieving high prediction accuracy under certain budget constraint. In automated medical diagnosis each medical test carries a monetary cost and the aim is to correctly diagnose any given patient based on a sequence of tests while keeping an expected budget. In search engines the speed of producing search results is often just as important as the quality of the results. Given a budget in terms of CPU time, the goal is to have a ranking algorithm that produces accurate ranking results within the budget.

Learning under resource constraints departs from the traditional machine learning setting and introduces new exciting challenges. For instance, features are accompanied by costs (e.g. extraction time in search engines or true monetary values in medical diagnosis) and their amortized sum is constrained at test-time.

This introduces a new trade-off between prediction accuracy and prediction cost. Studying this tradeoff is an inherent challenge that needs to be investigated in a principled fashion in order to invent practically relevant machine learning algorithms. 


Publications:

Selective Classification Based on One-Sided Prediction, AISTATS 2021

Budget Learning by Bracketing, AISTATS 2020

Adaptive Classification for Prediction under a Budget, NIPS 2017

Adaptive Neural Networks for Fast Test-Time Prediction, ICML 2017

Sequential Dynamic Decision Making with Deep Neural Nets on a Test-Time Budget, arxiv preprint

Structured Prediction on a Budget, AAAI 2017

Pruning Random Forests for Prediction on a Budget, NIPS 2016

Efficient Learning by Directed Acyclic Graph For Resource Constrained Prediction, NIPS 2015

Cheap Bandits, ICML 2015

Feature-Budgeted Random Forest, ICML 2015

Model Selection by Linear Programming, ECCV 2014

An LP for Sequential Learning Under Budgets, AISTATS 2014

Fast Margin-based Cost-Sensitive Classification, ICASSP 2014

Multistage Learning under Budget Constraints, AISTATS 2013 (Oral)

Multi Stage Classifier Design, ACML 2012