Data pruning is observed to substantially reduce the computation and memory costs of model training. Previous studies have primarily focused on constructing a series of coresets with representative samples by leveraging predefined rules for evaluating sample importance. Learning dynamics and selection bias, however, are rarely being considered. In this letter, a novel Sample-wise Dynamic Probabilistic Pruning (SwDPP) method is proposed for efficient training. Specifically, instead of hard-pruning the samples that are considered easy or well-learned, we formulate the pruning process as a probabilistic sampling problem. This is achieved by a carefully-designed soft-selection mechanism, which constantly expresses learning dynamics and relaxes selection bias. Moreover, to alleviate the accuracy drop under high pruning rates, we introduce a probabilistic Mixup strategy for information diversity maintenance. Extensive experiments conducted on CIFAR-10, CIFAR-100 and Tiny-ImageNet show that, the proposed SwDPP outperforms current state-of-the-art methods across various pruning settings. Notably, on CIFAR-10 and CIFAR-100, SwDPP achieves lossless training acceleration using only 70% of the data per epoch.

The performance of image captioning has been significantly improved recently through deep neural network architectures combining with attention mechanisms and reinforcement learning optimization. Exploring visual relationships and interactions between different objects appearing in the image, however, is far from being investigated. In this paper, we present a novel approach that combines scene graphs with Transformer, which we call SGT, to explicitly encode available visual relationships between detected objects. Specifically, we pretrain an scene graph generation model to predict graph representations for images. After that, for each graph node, a Graph Convolutional Network (GCN) is employed to acquire relationship knowledge by aggregating the information of its local neighbors. As we train the captioning model, we feed the potential relation-aware information into the Transformer to generate descriptive sentence. Experiments on the MS COCO dataset validate the superiority of our SGT model, which can realize state-of-the-art results in terms of all the standard evaluation metrics.

Most existing image captioning methods use only the visual information of the image to guide the generation of captions, lack the guidance of effective scene semantic information, and the current visual attention mechanism cannot adjust the focus intensity on the image. In this article, we first propose an improved visual attention model. At each timestep, we calculated the focus intensity coefficient of the attention mechanism through the context information of the model, then automatically adjusted the focus intensity of the attention mechanism through the coefficient to extract more accurate visual information. In addition, we represented the scene semantic knowledge of the image through topic words related to the image scene, then added them to the language model. We used the attention mechanism to determine the visual information and scene semantic information that the model pays attention to at each timestep and combined them to enable the model to generate more accurate and scene-specific captions. Finally, we evaluated our model on Microsoft COCO (MSCOCO) and Flickr30k standard datasets. The experimental results show that our approach generates more accurate captions and outperforms many recent advanced models in various evaluation metrics.

Automatically generating a human-like description for a given image is a potential research in artificial intelligence, which has attracted a great of attention recently. Most of the existing attention methods explore the mapping relationships between words in sentence and regions in image, such unpredictable matching manner sometimes causes inharmonious alignments that may reduce the quality of generated captions. In this paper, we make our efforts to reason about more accurate and meaningful captions. We first propose word attention to improve the correctness of visual attention when generating sequential descriptions word-by-word. The special word attention emphasizes on word importance when focusing on different regions of the input image, and makes full use of the internal annotation knowledge to assist the calculation of visual attention. Then, in order to reveal those incomprehensible intentions that cannot be expressed straightforwardly by machines, we inject external knowledge extracted from knowledge graph into the encoder-decoder framework to facilitate meaningful captioning. We validate our model on two freely available captioning benchmarks: Microsoft COCO dataset and Flickr30k dataset. The results demonstrate that our approach achieves state-of-the-art performance and outperforms many of the existing approaches.

Due to the deteriorated quality of feature in the propagation process of the neural network, it may be hard for traditional detector to identify a small object by just utilizing information within one region proposal. To overcome the limitation of the traditional object detector, we proposed a graph based relation mining network, to capture the relation information from labels and images. The semantic relation network is proposed to mine the global semantic relation in labels, and the spatial relation network is proposed to capture the local spatial relation in images. The feature representation is further improved by aggregating the outputs of the two networks. Instead of directly disseminating visual features in the network, the relation mining network explores more advanced feature information. Experiments on the PASCAL VOC and MS COCO datasets demonstrate that key relation information significantly improve the performance of object detection with better ability to detect small objects and reasonable bounding box. The results on COCO dataset demonstrate our method can detect objects robustly, increasing the detection performance of small objects from average precision and average recall by 4.7% and 7.6% respectively in performance relative to Faster R-CNN.

Most object detection methods focus only on the local information near the region proposal and ignore the object's global semantic relation and local spatial relation information, resulting in limited performance. To capture and explore these important relations, we propose a detection method based on a graph convolutional network (GCN). Two independent relation graph networks are used to obtain the global semantic information of the object in labels and the local spatial information in images. Semantic relation networks can implicitly acquire global knowledge, and by constructing a directed graph on the dataset, each node is represented by the word embedding of labels and then sent to the GCN to obtain high-level semantic representation. The spatial relation network encodes the relation by the positional relation module and the visual connection module, and enriches the object features through local key information from objects. The feature representation is further improved by aggregating the outputs of the two networks. Instead of directly disseminating visual features in the network, the dual-graph network explores more advanced feature information, giving the detector the ability to obtain key relations in labels and region proposals. Experiments on the PASCAL VOC and MS COCO datasets demonstrate that key relation information significantly improve the performance of detection with better ability to detect small objects and reasonable boduning box. The results on COCO dataset demonstrate our method obtains around 32.3% improvement on AP in terms of small objects.