Calibrating Generative Models
This addresses calibration issues in generative models, which is an incremental improvement for applications requiring accurate probability estimates.
The paper tackled the problem of miscalibration in generative models by framing it as a constrained optimization problem and introducing two surrogate objectives for fine-tuning, resulting in substantial reductions in calibration error across models with up to one billion parameters in applications like protein design, image generation, and language modeling.
Generative models frequently suffer miscalibration, wherein class probabilities and other statistics of the sampling distribution deviate from desired values. We frame calibration as a constrained optimization problem and seek the closest model in Kullback-Leibler divergence satisfying calibration constraints. To address the intractability of imposing these constraints exactly, we introduce two surrogate objectives for fine-tuning: (1) the relax loss, which replaces the constraint with a miscalibration penalty, and (2) the reward loss, which converts calibration into a reward fine-tuning problem. We demonstrate that these approaches substantially reduce calibration error across hundreds of simultaneous constraints and models with up to one billion parameters, spanning applications in protein design, image generation, and language modeling.