Adversarial training for predictive tasks: theoretical analysis and limitations in the deterministic case

TL;DR

This paper investigates the effectiveness of adversarial training for deterministic sequence prediction tasks, providing a theoretical explanation for its limitations and proposing an alternative adversarial content loss method.

Contribution

It offers a theoretical analysis explaining why CGANs may not improve deterministic sequence predictions and introduces an adversarial content loss approach for better results.

Findings

CGANs do not significantly improve deterministic sequence predictions compared to $L_p$ loss.

A theoretical framework explains the limitations of adversarial training in deterministic cases.

An adversarial content loss method yields better results on geophysical data.

Abstract

To train a deep neural network to mimic the outcomes of processing sequences, a version of Conditional Generalized Adversarial Network (CGAN) can be used. It has been observed by others that CGAN can help to improve the results even for deterministic sequences, where only one output is associated with the processing of a given input. Surprisingly, our CGAN-based tests on deterministic geophysical processing sequences did not produce a real improvement compared to the use of an $L_p$ loss; we here propose a first theoretical explanation why. Our analysis goes from the non-deterministic case to the deterministic one. It led us to develop an adversarial way to train a content loss that gave better results on our data.