A real-world text corpus sometimes comprises not only text documents but also semantic links between them (e.g., academic papers in a bibliographic network are linked by citations and co-authorships). Text documents and semantic connections form a text-rich network, which empowers a wide range of downstream tasks such as classification and retrieval. However, pretraining methods for such structures are still lacking, making it difficult to build one generic model that can be adapted to various tasks on text-rich networks. Current pretraining objectives, such as masked language modeling, purely model texts and do not take inter-document structure information into consideration. To this end, we propose our PretrAining on TexT-Rich NetwOrk framework Patton. Patton includes two pretraining strategies: network-contextualized masked language modeling and masked node prediction, to capture the inherent dependency between textual attributes and network structure. We conduct experiments on four downstream tasks in five datasets from both academic and e-commerce domains, where Patton outperforms baselines significantly and consistently.
Video colorization task has recently attracted wide attention. Recent methods mainly work on the temporal consistency in adjacent frames or frames with small interval. However, it still faces severe challenge of the inconsistency between frames with large interval.To address this issue, we propose a novel video colorization framework, which combines semantic correspondence into automatic video colorization to keep long-range consistency. Firstly, a reference colorization network is designed to automatically colorize the first frame of each video, obtaining a reference image to supervise the following whole colorization process. Such automatically colorized reference image can not only avoid labor-intensive and time-consuming manual selection, but also enhance the similarity between reference and grayscale images. Afterwards, a semantic correspondence network and an image colorization network are introduced to colorize a series of the remaining frames with the help of the reference. Each frame is supervised by both the reference image and the immediately colorized preceding frame to improve both short-range and long-range temporal consistency. Extensive experiments demonstrate that our method outperforms other methods in maintaining temporal consistency both qualitatively and quantitatively. In the NTIRE 2023 Video Colorization Challenge, our method ranks at the 3rd place in Color Distribution Consistency (CDC) Optimization track.
The Segment Anything Model (SAM) has recently gained popularity in the field of image segmentation. Thanks to its impressive capabilities in all-round segmentation tasks and its prompt-based interface, SAM has sparked intensive discussion within the community. It is even said by many prestigious experts that image segmentation task has been "finished" by SAM. However, medical image segmentation, although an important branch of the image segmentation family, seems not to be included in the scope of Segmenting "Anything". Many individual experiments and recent studies have shown that SAM performs subpar in medical image segmentation. A natural question is how to find the missing piece of the puzzle to extend the strong segmentation capability of SAM to medical image segmentation. In this paper, instead of fine-tuning the SAM model, we propose Med SAM Adapter, which integrates the medical specific domain knowledge to the segmentation model, by a simple yet effective adaptation technique. Although this work is still one of a few to transfer the popular NLP technique Adapter to computer vision cases, this simple implementation shows surprisingly good performance on medical image segmentation. A medical image adapted SAM, which we have dubbed Medical SAM Adapter (MSA), shows superior performance on 19 medical image segmentation tasks with various image modalities including CT, MRI, ultrasound image, fundus image, and dermoscopic images. MSA outperforms a wide range of state-of-the-art (SOTA) medical image segmentation methods, such as nnUNet, TransUNet, UNetr, MedSegDiff, and also outperforms the fully fine-turned MedSAM with a considerable performance gap. Code will be released at: https://github.com/WuJunde/Medical-SAM-Adapter.
Weakly supervised semantic segmentation (WSSS) based on image-level labels is challenging since it is hard to obtain complete semantic regions. To address this issue, we propose a self-training method that utilizes fused multi-scale class-aware attention maps. Our observation is that attention maps of different scales contain rich complementary information, especially for large and small objects. Therefore, we collect information from attention maps of different scales and obtain multi-scale attention maps. We then apply denoising and reactivation strategies to enhance the potential regions and reduce noisy areas. Finally, we use the refined attention maps to retrain the network. Experiments showthat our method enables the model to extract rich semantic information from multi-scale images and achieves 72.4% mIou scores on both the PASCAL VOC 2012 validation and test sets. The code is available at https://bupt-ai-cz.github.io/SMAF.
Accurate load forecasting is critical for reliable and efficient planning and operation of electric power grids. In this paper, we propose a unifying deep learning framework for load forecasting, which includes time-varying feature weighting, hierarchical temporal attention, and feature-reinforced error correction. Our framework adopts a modular design with good generalization capability. First, the feature-weighting mechanism assigns input features with temporal weights. Second, a recurrent encoder-decoder structure with hierarchical attention is developed as a load predictor. The hierarchical attention enables a similar day selection, which re-evaluates the importance of historical information at each time step. Third, we develop an error correction module that explores the errors and learned feature hidden information to further improve the model's forecasting performance. Experimental results demonstrate that our proposed framework outperforms existing methods on two public datasets and performance metrics, with the feature weighting mechanism and error correction module being critical to achieving superior performance. Our framework provides an effective solution to the electric load forecasting problem, which can be further adapted to many other forecasting tasks.
We propose to Transform Scene Graphs (TSG) into more descriptive captions. In TSG, we apply multi-head attention (MHA) to design the Graph Neural Network (GNN) for embedding scene graphs. After embedding, different graph embeddings contain diverse specific knowledge for generating the words with different part-of-speech, e.g., object/attribute embedding is good for generating nouns/adjectives. Motivated by this, we design a Mixture-of-Expert (MOE)-based decoder, where each expert is built on MHA, for discriminating the graph embeddings to generate different kinds of words. Since both the encoder and decoder are built based on the MHA, as a result, we construct a homogeneous encoder-decoder unlike the previous heterogeneous ones which usually apply Fully-Connected-based GNN and LSTM-based decoder. The homogeneous architecture enables us to unify the training configuration of the whole model instead of specifying different training strategies for diverse sub-networks as in the heterogeneous pipeline, which releases the training difficulty. Extensive experiments on the MS-COCO captioning benchmark validate the effectiveness of our TSG. The code is in: https://anonymous.4open.science/r/ACL23_TSG.
The retrieval model is an indispensable component for real-world knowledge-intensive tasks, e.g., open-domain question answering (ODQA). As separate retrieval skills are annotated for different datasets, recent work focuses on customized methods, limiting the model transferability and scalability. In this work, we propose a modular retriever where individual modules correspond to key skills that can be reused across datasets. Our approach supports flexible skill configurations based on the target domain to boost performance. To mitigate task interference, we design a novel modularization parameterization inspired by sparse Transformer. We demonstrate that our model can benefit from self-supervised pretraining on Wikipedia and fine-tuning using multiple ODQA datasets, both in a multi-task fashion. Our approach outperforms recent self-supervised retrievers in zero-shot evaluations and achieves state-of-the-art fine-tuned retrieval performance on NQ, HotpotQA and OTT-QA.
Sharpness-aware minimization (SAM), which searches for flat minima by min-max optimization, has been shown to be useful in improving model generalization. However, since each SAM update requires computing two gradients, its computational cost and training time are both doubled compared to standard empirical risk minimization (ERM). Recent state-of-the-arts reduce the fraction of SAM updates and thus accelerate SAM by switching between SAM and ERM updates randomly or periodically. In this paper, we design an adaptive policy to employ SAM based on the loss landscape geometry. Two efficient algorithms, AE-SAM and AE-LookSAM, are proposed. We theoretically show that AE-SAM has the same convergence rate as SAM. Experimental results on various datasets and architectures demonstrate the efficiency and effectiveness of the adaptive policy.
Recently, a number of approaches to train speech models by incorpo-rating text into end-to-end models have been developed, with Mae-stro advancing state-of-the-art automatic speech recognition (ASR)and Speech Translation (ST) performance. In this paper, we expandour understanding of the resulting shared speech-text representationswith two types of analyses. First we examine the limits of speech-free domain adaptation, finding that a corpus-specific duration modelfor speech-text alignment is the most important component for learn-ing a shared speech-text representation. Second, we inspect the sim-ilarities between activations of unimodal (speech or text) encodersas compared to the activations of a shared encoder. We find that theshared encoder learns a more compact and overlapping speech-textrepresentation than the uni-modal encoders. We hypothesize that thispartially explains the effectiveness of the Maestro shared speech-textrepresentations.