Socratic Learning: Correcting Misspecified Generative Models using Discriminative Models

A challenge in training discriminative models like neural networks is obtaining enough labeled training data. Recent approaches have leveraged generative models to denoise weak supervision sources that a discriminative model can learn from. These generative models directly encode the users' background knowledge. Therefore, these models may be incompletely specified and fail to model latent classes in the data. We present Socratic learning to systematically correct such generative model misspecification by utilizing feedback from the discriminative model. We prove that under mild conditions, Socratic learning can recover features from the discriminator that informs the generative model about these latent classes. Experimentally, we show that without any hand-labeled data, the corrected generative model improves discriminative performance by up to 4.47 points and reduces error for an image classification task by 80% compared to a state-of-the-art weak supervision modeling technique.