Perspective Chapter: Cross-Level Unbiased Prior Knowledge-Guided Deep Representation Learning

The great success of deep representation learning largely depends on the large amount of data available for model training. However, labeling data that satisfies the assumption of independent and identical distribution brings unaffordable workloads, resulting in the connectionism-based data-driven learning paradigm encountering a development bottleneck. The essential reason is that the traditional deep representation learning paradigm aims to enable the model to perform supervised training on a specific data domain, so the model can only obtain the domain-biased discriminative information, and the insufficient supervised data directly incurs the model not to obtain sufficient discriminative information. To remedy such a deficiency, researchers focus on introducing unbiased prior knowledge to guide the training of deep representation learning models, to improve the model’s unbiased discriminative representation ability, thereby enhancing the performance of deep learning models on discriminative tasks.