Over the past two decades, deep learning (DL) has achieved unprecedented breakthroughs across diverse application domains spanning computer vision (CV) to natural language processing (NLP). However, despite significant advances in computational resources and algorithmic frameworks, the training of deep neural networks continues to present formidable challenges due to persistent issues of training inefficiency and inherent data distribution biases. Recent years have witnessed the emergence of hard sample mining (HSM) as a promising paradigm to mitigate training inefficiencies and enhance model robustness through representative sample selection. Although HSM is reshaping contemporary AI research, its critical role in enabling efficient and robust model training has not yet been systematically explored. This article presents a comprehensive survey of HSM methodologies by: 1) establishing unified definitions of hard samples through rigorous sample complexity quantification criteria; 2) proposing a systematic taxonomy of HSM approaches with in-depth technical analysis; and 3) identifying pivotal research frontiers in this evolving field. This survey not only consolidates the foundations of HSM but also provides a roadmap for advancing efficient, robust, and generalizable deep learning models.

Learning causality from large-scale text corpora is an important task with numerous applications—for example, in finance, biology, medicine, and scientific discovery. Prior studies have focused mainly on simple causality, which only includes one cause-effect pair. However, causality is notoriously difficult to understand and analyze because of multiple cause spans and their entangled interactions. To detect complex causality, we propose a self-paced contrastive learning model, namely N2NCause, to learn entangled interactions between multiple spans. Specifically, N2NCause introduces data enhancement operations to convert implicit expressions into explicit expressions with the most rational causal connectives for the synthesis of positive samples and to invert the directed connection between a cause-effect pair for the synthesis of negative samples. To learn the semantic dependency and causal direction of positive and negative samples, self-paced contrastive learning is proposed to learn the entangled interactions among spans, including the interaction direction and interaction field. We evaluated the performance of N2NCause in three cause-effect detection tasks. The experimental results show that, with the least data annotation efforts, N2NCause demonstrates competitive performance in detecting simple cause-effect relations, and it is superior to existing solutions for the detection of complex causality.

Training deep neural networks (DNNs) on largescale datasets is often inefficient with large computational needs and significant energy consumption. Although great efforts have been taken to optimize DNNs, few studies focused on the inefficiency caused by the data samples with less value for model training. In this article, we empirically demonstrate that sample complexity is important for model efficiency and selecting representative samples is constructive to the model efficiency. In particular, we propose hardness-aware dataset pruning method (γ-Razor) to select representative samples from large-scale datasets to remove the less valuable data samples for model training.γ -Razor is a two-stage framework that includes interclass sampling and intraclass sampling. First, we introduce the inverse self-paced learning strategy to learn hard samples and adjust their weights adaptively according to the inverse frequency of effective samples of each class. For intraclass sampling, hardness-aware cluster sampling algorithm is proposed to downsample easy samples within each class. To evaluate the performance of γ-Razor, we conducted extensive experiments on three large-scale datasets for image classification tasks. The experimental results show that models trained with the pruned datasets show competitive performances against their counterparts trained with the original large-scale datasets in terms of robustness and efficiency. Furthermore, models trained with the pruned datasets converge faster with lower energy consumption.

Self-supervised monocular depth estimation has attracted significant attention due to its broad applications in autonomous driving and robotics. Although significant performance improvements have been achieved by learning the relative distance of objects with the introduction of Self Query Layer (SQL), it struggles with zero-shot generalization due to the lack of geometric features and the fixed number of query sizes. To address these problems, we propose a diffusion-augmented self-supervised depth estimation framework, named DiffSQL, to learn geometric priors for feature augmentation. Additionally, we introduce a dynamic self-query layer that implicitly computes the relative distances between objects by adjusting the query size according to the feature distribution. Experimental results on the KITTI dataset show that DiffSQL outperforms SQLdepth by 1.03% in terms of AbsRel and 2.79% in terms of SqRel. Furthermore, our experiments demonstrate that DiffSQL is superior in zero-shot generalization.