Direction Concentration Learning: Enhancing Congruency in Machine Learning

One of the well-known challenges in computer vision tasks is visual diversity images, which could result an agreement or disagreement between learned knowledge and content exhibited by current observation. In this work, we first define such a concepts learning process as congruency. Formally, given particular task sufficiently large dataset, congruency issue occurs whereby task-specific semantics training data are highly varying. We propose Direction Concentration Learning (DCL) method to improve process, where enhancing influences convergence path be less circuitous. The experimental results show that proposed DCL generalizes state-of-the-art models optimizers, well improves performances saliency prediction task, continual classification task. Moreover, it helps mitigate catastrophic forgetting problem code publicly available at https://github.com/luoyan407/congruency.