Feature Equilibrium: An Adversarial Training Method to Improve Representation Learning

Abstract Over-fitting is a significant threat to the integrity and reliability of deep neural networks with generous parameters. One problem that model learned more specific features than general in training process. To solve problem, we propose an adversarial method assist strengthening representation learning. In this method, make classification as generator G introduce unsupervised discriminator D distinguish hidden feature from real images limit their spatial distance. Notably, will fall into trap perfect resulting gradient confrontation loss 0 after overtraining. avoid situation, train probability $$P_{c}$$ P c . Our proposed easy incorporate existing frameworks. It has been evaluated under various network architectures over different fields datasets. Experiments show low computational cost, outperforms benchmark by 1.5–2 points on For semantic segmentation VOC, our achieves 2.2 higher mAP.