Reproduce “Algorithmic Collective Action in Machine Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Moritz Hardt, Eric Mazumdar, Celestine Mendler-Dünner, Tijana Zrnic
This research from the Max Planck Institute, Caltech, and UC Berkeley establishes a theoretical framework and provides empirical evidence for "algorithmic collective action," demonstrating that surprisingly small collectives can significantly influence machine learning models by strategically modifying their data. It quantifies the "critical mass" required for such collectives to steer algorithms, showing effectiveness even for deep learning models on a real-world gig economy dataset.
Key insights
- Algorithmic collectives of exceedingly small fractional sizes can exert substantial control over learning algorithms, with effectiveness depending on factors like signal uniqueness and classifier optimality.
- More optimized models (achieved through more training epochs) are more susceptible to collective influence, requiring a lower critical mass for success.
- Collective strategies are robust to the precise placement of trigger tokens and can even steer nonconvex gradient-based learning if adaptively applied.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.