Earning Extra Performance From Restrictive Feedbacks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Optimization

Background:

• Machine learning models often require refinement for specific user needs.
• Traditional model tuning relies on access to target data, which is frequently unavailable.
• Existing methods struggle when only performance evaluations, not gradients, can be accessed.

Purpose of the Study:

• To formally define the Earning eXtra PerformancE from restriCTive feEDdbacks (EXPECTED) challenge.
• To develop algorithms for tuning machine learning models using limited, scalar feedback.
• To enable model providers to improve models without direct access to user data.

Main Methods:

• Characterizing model performance geometry through parameter distribution exploration.
• Developing a query-efficient, layerwise tuning algorithm for deep models.
• Utilizing scalar feedback (e.g., accuracy, usage rate) for model refinement.

Main Results:

• Proposed algorithms demonstrate efficacy and efficiency in theoretical analyses.
• Extensive experiments validate the approach across diverse applications.
• A robust solution is established for the EXPECTED problem.

Conclusions:

• The EXPECTED challenge addresses a critical gap in practical model tuning.
• The proposed methods enable effective model adaptation with restrictive feedback.
• This work provides a foundation for future research in feedback-driven model optimization.