Reproduce “Learning from Streaming Data when Users Choose” on OpenResearch

About this paper

Jinyan Su, Sarah Dean

Learning from Streaming Data when Users Choose by Jinyan Su and Sarah Dean from Cornell University proposes Multi-learner Streaming Gradient Descent (MSGD) for decentralized machine learning systems where user choices dynamically shape data distributions. The paper provides theoretical guarantees for MSGD's convergence to stationary points of a non-convex overall loss function and empirically demonstrates that user choice leads to emergent specialization, improving accuracy for specific user subpopulations but potentially decreasing it for the entire population.

Key insights

• The overall loss function, which quantifies user welfare considering their choices, is generally non-convex.
• Despite non-convexity and partial information, the gradient of this overall loss for a service depends only on that service's local loss gradient, enabling decentralized optimization without explicit coordination.
• User choice naturally leads to emergent specialization among services, where each service optimizes for its specific subpopulation of users.

Summary by alphaXiv.