← Research

CoT: Cooperative Training for Generative Modeling of Discrete Data

Authors

Sidi Lu, Lantao Yu, Siyuan Feng, Yaoming Zhu, and Weinan Zhang

Overview

Cooperative Training, or CoT, is a training framework for generative models over discrete sequences. It avoids the unstable REINFORCE signal commonly used to pass discriminator feedback through non-differentiable token sampling.

CoT jointly trains a generator and an auxiliary predictive mediator. This turns the adversarial min–max game into a cooperative maximization objective and gives the model an explicit estimate of Jensen–Shannon divergence.

Results

Across discrete sequence-generation experiments, CoT achieves superior or competitive sample quality and diversity while improving training stability. It also works without the maximum-likelihood pretraining stage required by many adversarial baselines.

Paper

PMLR

Status

ICML 2019 Short Oral.

Citation

Sidi Lu, Lantao Yu, Siyuan Feng, Yaoming Zhu, and Weinan Zhang. “CoT: Cooperative Training for Generative Modeling of Discrete Data.” Proceedings of the 36th International Conference on Machine Learning, PMLR 97:4164–4172, 2019.