Self-supervised learning (SSL) aims to learn meaningful representations from unlabeled data. Orthogonal Low-rank Embedding (OLE) shows promise for SSL by enhancing intra-class similarity in a low-rank subspace and promoting inter-class dissimilarity in a high-rank subspace, making it particularly suitable for multi-view learning tasks. However, directly applying OLE to SSL poses significant challenges: (1) the virtually infinite number of "classes" in SSL makes achieving the OLE objective impractical, leading to representational collapse; and (2) low-rank constraints may fail to distinguish between positively and negatively correlated features, further undermining learning. To address these issues, we propose SSOLE (Self-Supervised Orthogonal Low-rank Embedding), a novel framework that integrates OLE principles into SSL by (1) decoupling the low-rank and high-rank enforcement to align with SSL objectives; and (2) applying low-rank constraints to feature deviations from their mean, ensuring better alignment of positive pairs by accounting for the signs of cosine similarities. Our theoretical analysis and empirical results demonstrate that these adaptations are crucial to SSOLE’s effectiveness. Moreover, SSOLE achieves competitive performance across SSL benchmarks without relying on large batch sizes, memory banks, or dual-encoder architectures, making it an efficient and scalable solution for self-supervised tasks. Code is available at https://github.com/husthuaan/ssole.

Exploring the entity-level (i.e., objects in an image, words in a text) spatial relationship contributes to understanding multimedia content precisely. The ignorance of spatial information in previous works probably leads to misunderstandings of image contents. For instance, sentences `Boats are on the water' and `Boats are under the water' describe the same objects, but correspond to different sceneries. To this end, we utilize the relative position of objects to capture entity-level spatial relationships for image-text matching. Specifically, we fuse semantic and spatial relationships of image objects in a visual intra-modal relation module. The module performs promisingly to understand image contents and improve object representation learning. It contributes to capturing entity-level latent correspondence of image-text pairs. Then the query (text) plays a role of textual context to refine the interpretable alignments of image-text pairs in the inter-modal relation module. Our proposed method achieves state-of-the-art results on MSCOCO and Flickr30K datasets.

Recent neural models on image captioning usually take a encoder-decoder fashion, where the decoder predicts a single word at one step recently with the encoder providing information. The encoder is a pretrained CNN model typically. Thus the decoder, the input to it, and the output from it become the most important parts of a model. We propose a pipelined image captioning framework consisting of two cascaded agents. The former is named as "semantic adaptive agent" which generates the input to the decoder by consulting the information from the current decoding process, and the latter as "caption generating agent" which select a single word of the vocabulary as the output of the decoder by taking consideration of the input and the current states of the decoder. For the framework of two cascaded agents, we design a multi-stage training procedure to train the two agents with different objectives by fully utilizing reinforcement learning. In experiments, we conduct quantitative and qualitative analysis on MS COCO dataset and our results can significantly outperform baseline methods in terms of several evaluation metrics.

Attention mechanisms are widely used in current encoder/decoder frameworks of image captioning, where a weighted average on encoded vectors is generated at each time step to guide the caption decoding process. However, the decoder has little idea of whether or how well the attended vector and the given attention query are related, which could make the decoder give misled results. In this paper, we propose an Attention on Attention (AoA) module, which extends the conventional attention mechanisms to determine the relevance between attention results and queries. AoA first generates an information vector and an attention gate using the attention result and the current context, then adds another attention by applying element-wise multiplication to them and finally obtains the attended information, the expected useful knowledge. We apply AoA to both the encoder and the decoder of our image captioning model, which we name as AoA Network (AoANet). Experiments show that AoANet outperforms all previously published methods and achieves a new state-of-the-art performance of 129.8 CIDEr-D score on MS COCO Karpathy offline test split and 129.6 CIDEr-D (C40) score on the official online testing server. Code is available at https://github.com/husthuaan/AoANet.

Recent neural models for image captioning usually employ an encoder-decoder framework with an attention mechanism. However, the attention mechanism in such a framework aligns one single (attended) image feature vector to one caption word, assuming one-to-one mapping from source image regions and target caption words, which is never possible. In this paper, we propose a novel attention model, namely Adaptive Attention Time (AAT), to align the source and the target adaptively for image captioning. AAT allows the framework to learn how many attention steps to take to output a caption word at each decoding step. With AAT, an image region can be mapped to an arbitrary number of caption words while a caption word can also attend to an arbitrary number of image regions. AAT is deterministic and differentiable, and doesn't introduce any noise to the parameter gradients. In this paper, we empirically show that AAT improves over state-of-the-art methods on the task of image captioning. Code is available at https://github.com/husthuaan/AAT.