Summary
This paper, "Cheap Thrills: Effective Amortized Optimization Using Inexpensive Labels" by Khai Nguyen, Petros Ellinas, Anvita Bhagavathula, and Priya Donti (arXiv:2603.05495, cs.LG, math.OC), addresses scaling the solution of optimization and simulation problems via machine-learning surrogates that map problem parameters to solutions. Existing approaches—supervised learning and self-supervised learning with soft or hard feasibility enforcement—face trade-offs such as reliance on expensive, high-quality labels or difficult optimization landscapes. The authors propose a novel framework that first collects "cheap" imperfect labels, performs supervised pretraining on them, and then refines the model through self-supervised learning to improve overall performance. Their theoretical analysis and a merit-based criterion show that labeled data need only place the model within a basin of attraction for the refinement stage to succeed, reducing labeling cost requirements. Shared on zhichai.net as an automatically collected arXiv paper digest (2026-03-07).
Cheap Thrills: Effective Amortized Optimization Using Inexpensive Labels
Authors: Khai Nguyen, Petros Ellinas, Anvita Bhagavathula, Priya Donti
arXiv: 2603.05495
PDF: https://arxiv.org/pdf/2603.05495.pdf
Categories: cs.LG, math.OC
Overview
This paper proposes a new method for amortized optimization: using machine-learning surrogates to inexpensively map problem parameters to corresponding solutions, thereby scaling the solution of optimization and simulation problems.
Motivation
Commonly used approaches—including supervised learning and self-supervised learning with either soft or hard feasibility enforcement—face inherent challenges:
- Reliance on expensive, high-quality labels
- Difficult optimization landscapes
Proposed Framework
To address these trade-offs, the authors introduce a framework with three stages:
1. Collect "cheap" imperfect labels
2. Perform supervised pretraining on these labels
3. Refine the model through self-supervised learning to improve overall performance
Key Finding
Theoretical analysis and a merit-based criterion show that labeled data need only place the model within a basin of attraction—expensive, high-quality labels are not strictly required for effective refinement.
---
*Auto-collected on 2026-03-07*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177168728