Machine learning agents improve semi supervised learning with unlabeled data
AgentSSL: Can MLE Agents Leverage Unlabeled Data?
Artificial IntelligenceMachine Learning
Summary
When teaching computers to learn, getting enough labeled examples can be very hard, while unlabeled data is often plentiful. The authors look at how machine learning agents can create programs that use both labeled and unlabeled data together, a method called semi-supervised learning. These agents were able to make better programs than human experts in many cases, but picking the best program from many candidates remains a big challenge. To study this, the authors created a large collection of programs to help test different ways of choosing the best one, and tried their approach on real-world ecology tasks.
What this means in practice
- •For machine learning engineers: Automatically generate semi-supervised learning pipelines that utilize unlabeled data when labeled data is limited.
- •For ecology data teams: Deploy agent-designed semi-supervised learning strategies to improve species detection in ecological datasets with scarce labels.
Authors
Ethan Lin, Akanksha Sarkar, Ziang Liu, Kristin Branson, Jennifer J. Sun
Abstract
Applying machine learning systems to new domains introduces bottlenecks in both labeled data acquisition and manual pipeline engineering. While recent MLE agents show promise in automating pipeline design, they have been studied almost exclusively in fully supervised settings, leaving them untested in realistic regimes where labeled data is scarce but unlabeled data is abundant. In this work, we introduce Agentic Semi-Supervised Learning (AgentSSL), framing SSL as a program search problem and studying how agents navigate this space under limited supervision. We find that even under weak feedback and matched compute, agents can synthesize code that combines and adapts established SSL techniques, exceeding the strongest human-engineered baselines in 8 of 12 settings by up to +21.8%. However, our results identify program selection as a critical bottleneck in AgentSSL, where performance drops by as much as 47% when selecting based on the agent's feedback metric. To systematically study this bottleneck, we introduce ProgramSelectBench, a large-scale benchmark with a combined total of 16,977 agent-generated and human-engineered SSL programs, and evaluate both unsupervised and active program selection methods. We further evaluate AgentSSL on real-world ecology detection tasks, highlighting the opportunities and challenges of transferring agent-discovered strategies beyond controlled benchmarks.