Auto-Encoder (AE)-based deep subspace clustering (DSC) methods have achieved impressive performance due to the powerful representation extracted using deep neural networks while prioritizing categorical separability. However, self-reconstruction loss of an AE ignores rich useful relation information and might lead to indiscriminative representation, which inevitably degrades the clustering performance. It is also challenging to learn high-level similarity without feeding semantic labels. Another unsolved problem facing DSC is the huge memory cost due to $n\times n$ similarity matrix, which is incurred by the self-expression layer between an encoder and decoder. To tackle these problems, we use pairwise similarity to weigh the reconstruction loss to capture local structure information, while a similarity is learned by the self-expression layer. Pseudo-graphs and pseudo-labels, which allow benefiting from uncertain knowledge acquired during network training, are further employed to supervise similarity learning. Joint learning and iterative training facilitate to obtain an overall optimal solution. Extensive experiments on benchmark datasets demonstrate the superiority of our approach. By combining with the $k$-nearest neighbors algorithm, we further show that our method can address the large-scale and out-of-sample problems.
Transducer-based models, such as RNN-Transducer and transformer-transducer, have achieved great success in speech recognition. A typical transducer model decodes the output sequence conditioned on the current acoustic state and previously predicted tokens step by step. Statistically, The number of blank tokens in the prediction results accounts for nearly 90\% of all tokens. It takes a lot of computation and time to predict the blank tokens, but only the non-blank tokens will appear in the final output sequence. Therefore, we propose a method named fast-skip regularization, which tries to align the blank position predicted by a transducer with that predicted by a CTC model. During the inference, the transducer model can predict the blank tokens in advance by a simple CTC project layer without many complicated forward calculations of the transducer decoder and then skip them, which will reduce the computation and improve the inference speed greatly. All experiments are conducted on a public Chinese mandarin dataset AISHELL-1. The results show that the fast-skip regularization can indeed help the transducer model learn the blank position alignments. Besides, the inference with fast-skip can be speeded up nearly 4 times with only a little performance degradation.
Fully exploiting ad-hoc microphone networks for distant speech recognition is still an open issue. Empirical evidence shows that being able to select the best microphone leads to significant improvements in recognition without any additional effort on front-end processing. Current channel selection techniques either rely on signal, decoder or posterior-based features. Signal-based features are inexpensive to compute but do not always correlate with recognition performance. Instead decoder and posterior-based features exhibit better correlation but require substantial computational resources. In this work, we tackle the channel selection problem by proposing MicRank, a learning to rank framework where a neural network is trained to rank the available channels using directly the recognition performance on the training set. The proposed approach is agnostic with respect to the array geometry and type of recognition back-end. We investigate different learning to rank strategies using a synthetic dataset developed on purpose and the CHiME-6 data. Results show that the proposed approach is able to considerably improve over previous selection techniques, reaching comparable and in some instances better performance than oracle signal-based measures.
Mandarin-English code-switching (CS) is frequently used among East and Southeast Asian people. However, the intra-sentence language switching of the two very different languages makes recognizing CS speech challenging. Meanwhile, the recent successful non-autoregressive (NAR) ASR models remove the need for left-to-right beam decoding in autoregressive (AR) models and achieved outstanding performance and fast inference speed, but it has not been applied to Mandarin-English CS speech recognition. This paper takes advantage of the Mask-CTC NAR ASR framework to tackle the CS speech recognition issue. We further propose to change the Mandarin output target of the encoder to Pinyin for faster encoder training and introduce the Pinyin-to-Mandarin decoder to learn contextualized information. Moreover, we use word embedding label smoothing to regularize the decoder with contextualized information and projection matrix regularization to bridge that gap between the encoder and decoder. We evaluate these methods on the SEAME corpus and achieved exciting results.
How can we effectively inform content selection in Transformer-based abstractive summarization models? In this work, we present a simple-yet-effective attention head masking technique, which is applied on encoder-decoder attentions to pinpoint salient content at inference time. Using attention head masking, we are able to reveal the relation between encoder-decoder attentions and content selection behaviors of summarization models. We then demonstrate its effectiveness on three document summarization datasets based on both in-domain and cross-domain settings. Importantly, our models outperform prior state-of-the-art models on CNN/Daily Mail and New York Times datasets. Moreover, our inference-time masking technique is also data-efficient, requiring only 20% of the training samples to outperform BART fine-tuned on the full CNN/DailyMail dataset.
This paper presents our recent effort on end-to-end speaker-attributed automatic speech recognition, which jointly performs speaker counting, speech recognition and speaker identification for monaural multi-talker audio. Firstly, we thoroughly update the model architecture that was previously designed based on a long short-term memory (LSTM)-based attention encoder decoder by applying transformer architectures. Secondly, we propose a speaker deduplication mechanism to reduce speaker identification errors in highly overlapped regions. Experimental results on the LibriSpeechMix dataset shows that the transformer-based architecture is especially good at counting the speakers and that the proposed model reduces the speaker-attributed word error rate by 47% over the LSTM-based baseline. Furthermore, for the LibriCSS dataset, which consists of real recordings of overlapped speech, the proposed model achieves concatenated minimum-permutation word error rates of 11.9% and 16.3% with and without target speaker profiles, respectively, both of which are the state-of-the-art results for LibriCSS with the monaural setting.
The quadratic computational and memory complexities of large Transformers have limited their scalability for long document summarization. In this paper, we propose Hepos, a novel efficient encoder-decoder attention with head-wise positional strides to effectively pinpoint salient information from the source. We further conduct a systematic study of existing efficient self-attentions. Combined with Hepos, we are able to process ten times more tokens than existing models that use full attentions. For evaluation, we present a new dataset, GovReport, with significantly longer documents and summaries. Results show that our models produce significantly higher ROUGE scores than competitive comparisons, including new state-of-the-art results on PubMed. Human evaluation also shows that our models generate more informative summaries with fewer unfaithful errors.
Transformer architectures show spectacular performance on NLP tasks and have recently also been used for tasks such as image completion or image classification. Here we propose to use a sequential image representation, where each prefix of the complete sequence describes the whole image at reduced resolution. Using such Fourier Domain Encodings (FDEs), an auto-regressive image completion task is equivalent to predicting a higher resolution output given a low-resolution input. Additionally, we show that an encoder-decoder setup can be used to query arbitrary Fourier coefficients given a set of Fourier domain observations. We demonstrate the practicality of this approach in the context of computed tomography (CT) image reconstruction. In summary, we show that Fourier Image Transformer (FIT) can be used to solve relevant image analysis tasks in Fourier space, a domain inherently inaccessible to convolutional architectures.
This paper introduces the task of few-shot common action localization in time and space. Given a few trimmed support videos containing the same but unknown action, we strive for spatio-temporal localization of that action in a long untrimmed query video. We do not require any class labels, interval bounds, or bounding boxes. To address this challenging task, we introduce a novel few-shot transformer architecture with a dedicated encoder-decoder structure optimized for joint commonality learning and localization prediction, without the need for proposals. Experiments on our reorganizations of the AVA and UCF101-24 datasets show the effectiveness of our approach for few-shot common action localization, even when the support videos are noisy. Although we are not specifically designed for common localization in time only, we also compare favorably against the few-shot and one-shot state-of-the-art in this setting. Lastly, we demonstrate that the few-shot transformer is easily extended to common action localization per pixel.
Retinal blood vessel can assist doctors in diagnosis of eye-related diseases such as diabetes and hypertension, and its segmentation is particularly important for automatic retinal image analysis. However, it is challenging to segment these vessels structures, especially the thin capillaries from the color retinal image due to low contrast and ambiguousness. In this paper, we propose pyramid U-Net for accurate retinal vessel segmentation. In pyramid U-Net, the proposed pyramid-scale aggregation blocks (PSABs) are employed in both the encoder and decoder to aggregate features at higher, current and lower levels. In this way, coarse-to-fine context information is shared and aggregated in each block thus to improve the location of capillaries. To further improve performance, two optimizations including pyramid inputs enhancement and deep pyramid supervision are applied to PSABs in the encoder and decoder, respectively. For PSABs in the encoder, scaled input images are added as extra inputs. While for PSABs in the decoder, scaled intermediate outputs are supervised by the scaled segmentation labels. Extensive evaluations show that our pyramid U-Net outperforms the current state-of-the-art methods on the public DRIVE and CHASE-DB1 datasets.
Crowd counting aims to predict the number of people and generate the density map in the image. There are many challenges, including varying head scales, the diversity of crowd distribution across images and cluttered backgrounds. In this paper, we propose a multi-scale context aggregation network (MSCANet) based on single-column encoder-decoder architecture for crowd counting, which consists of an encoder based on a dense context-aware module (DCAM) and a hierarchical attention-guided decoder. To handle the issue of scale variation, we construct the DCAM to aggregate multi-scale contextual information by densely connecting the dilated convolution with varying receptive fields. The proposed DCAM can capture rich contextual information of crowd areas due to its long-range receptive fields and dense scale sampling. Moreover, to suppress the background noise and generate a high-quality density map, we adopt a hierarchical attention-guided mechanism in the decoder. This helps to integrate more useful spatial information from shallow feature maps of the encoder by introducing multiple supervision based on semantic attention module (SAM). Extensive experiments demonstrate that the proposed approach achieves better performance than other similar state-of-the-art methods on three challenging benchmark datasets for crowd counting. The code is available at https://github.com/KingMV/MSCANet
Currently, a growing number of mature natural language processing applications make people's life more convenient. Such applications are built by source code - the language in software engineering. However, the applications for understanding source code language to ease the software engineering process are under-researched. Simultaneously, the transformer model, especially its combination with transfer learning, has been proven to be a powerful technique for natural language processing tasks. These breakthroughs point out a promising direction for process source code and crack software engineering tasks. This paper describes CodeTrans - an encoder-decoder transformer model for tasks in the software engineering domain, that explores the effectiveness of encoder-decoder transformer models for six software engineering tasks, including thirteen sub-tasks. Moreover, we have investigated the effect of different training strategies, including single-task learning, transfer learning, multi-task learning, and multi-task learning with fine-tuning. CodeTrans outperforms the state-of-the-art models on all the tasks. To expedite future works in the software engineering domain, we have published our pre-trained models of CodeTrans. https://github.com/agemagician/CodeTrans
In this paper, we describe the work that we have done to participate in Task1 of the ConferencingSpeech2021 challenge. This task set a goal to develop the solution for multi-channel speech enhancement in a real-time manner. We propose a novel system for streaming speech enhancement. We employ Wave-U-Net architecture with temporal convolutions in encoder and decoder. We incorporate self-attention in the decoder to apply attention mask retrieved from skip-connection on features from down-blocks. We explore history cache mechanisms that work like hidden states in recurrent networks and implemented them in proposal solution. It helps us to run an inference with chunks length 40ms and Real-Time Factor 0.4 with the same precision.
Relation extraction from text is an important task for automatic knowledge base population. In this thesis, we first propose a syntax-focused multi-factor attention network model for finding the relation between two entities. Next, we propose two joint entity and relation extraction frameworks based on encoder-decoder architecture. Finally, we propose a hierarchical entity graph convolutional network for relation extraction across documents.
How to generate summaries of different styles without requiring corpora in the target styles, or training separate models? We present two novel methods that can be deployed during summary decoding on any pre-trained Transformer-based summarization model. (1) Decoder state adjustment instantly modifies decoder final states with externally trained style scorers, to iteratively refine the output against a target style. (2) Word unit prediction constrains the word usage to impose strong lexical control during generation. In experiments of summarizing with simplicity control, automatic evaluation and human judges both find our models producing outputs in simpler languages while still informative. We also generate news headlines with various ideological leanings, which can be distinguished by humans with a reasonable probability.
The autoregressive (AR) models, such as attention-based encoder-decoder models and RNN-Transducer, have achieved great success in speech recognition. They predict the output sequence conditioned on the previous tokens and acoustic encoded states, which is inefficient on GPUs. The non-autoregressive (NAR) models can get rid of the temporal dependency between the output tokens and predict the entire output tokens in at least one step. However, the NAR model still faces two major problems. On the one hand, there is still a great gap in performance between the NAR models and the advanced AR models. On the other hand, it's difficult for most of the NAR models to train and converge. To address these two problems, we propose a new model named the two-step non-autoregressive transformer(TSNAT), which improves the performance and accelerating the convergence of the NAR model by learning prior knowledge from a parameters-sharing AR model. Furthermore, we introduce the two-stage method into the inference process, which improves the model performance greatly. All the experiments are conducted on a public Chinese mandarin dataset ASIEHLL-1. The results show that the TSNAT can achieve a competitive performance with the AR model and outperform many complicated NAR models.
Depth information has been proved beneficial in RGB-D salient object detection (SOD). However, depth maps obtained often suffer from low quality and inaccuracy. Most existing RGB-D SOD models have no cross-modal interactions or only have unidirectional interactions from depth to RGB in their encoder stages, which may lead to inaccurate encoder features when facing low quality depth. To address this limitation, we propose to conduct progressive bi-directional interactions as early in the encoder stage, yielding a novel bi-directional transfer-and-selection network named BTS-Net, which adopts a set of bi-directional transfer-and-selection (BTS) modules to purify features during encoding. Based on the resulting robust encoder features, we also design an effective light-weight group decoder to achieve accurate final saliency prediction. Comprehensive experiments on six widely used datasets demonstrate that BTS-Net surpasses 16 latest state-of-the-art approaches in terms of four key metrics.
How to effectively represent camera pose is an essential problem in 3D computer vision, especially in tasks such as camera pose regression and novel view synthesis. Traditionally, 3D position of the camera is represented by Cartesian coordinate and the orientation is represented by Euler angle or quaternions. These representations are manually designed, which may not be the most effective representation for downstream tasks. In this work, we propose an approach to learn neural representations of camera poses and 3D scenes, coupled with neural representations of local camera movements. Specifically, the camera pose and 3D scene are represented as vectors and the local camera movement is represented as a matrix operating on the vector of the camera pose. We demonstrate that the camera movement can further be parametrized by a matrix Lie algebra that underlies a rotation system in the neural space. The vector representations are then concatenated and generate the posed 2D image through a decoder network. The model is learned from only posed 2D images and corresponding camera poses, without access to depths or shapes. We conduct extensive experiments on synthetic and real datasets. The results show that compared with other camera pose representations, our learned representation is more robust to noise in novel view synthesis and more effective in camera pose regression.
The recently proposed end-to-end transformer detectors, such as DETR and Deformable DETR, have a cascade structure of stacking 6 decoder layers to update object queries iteratively, without which their performance degrades seriously. In this paper, we investigate that the random initialization of object containers, which include object queries and reference points, is mainly responsible for the requirement of multiple iterations. Based on our findings, we propose Efficient DETR, a simple and efficient pipeline for end-to-end object detection. By taking advantage of both dense detection and sparse set detection, Efficient DETR leverages dense prior to initialize the object containers and brings the gap of the 1-decoder structure and 6-decoder structure. Experiments conducted on MS COCO show that our method, with only 3 encoder layers and 1 decoder layer, achieves competitive performance with state-of-the-art object detection methods. Efficient DETR is also robust in crowded scenes. It outperforms modern detectors on CrowdHuman dataset by a large margin.
The accurate prognosis of Glioblastoma Multiforme (GBM) plays an essential role in planning correlated surgeries and treatments. The conventional models of survival prediction rely on radiomic features using magnetic resonance imaging (MRI). In this paper, we propose a radiogenomic overall survival (OS) prediction approach by incorporating gene expression data with radiomic features such as shape, geometry, and clinical information. We exploit TCGA (The Cancer Genomic Atlas) dataset and synthesize the missing MRI modalities using a fully convolutional network (FCN) in a conditional Generative Adversarial Network (cGAN). Meanwhile, the same FCN architecture enables the tumor segmentation from the available and the synthesized MRI modalities. The proposed FCN architecture comprises octave convolution (OctConv) and a novel decoder, with skip connections in spatial and channel squeeze & excitation (skip-scSE) block. The OctConv can process low and high-frequency features individually and improve model efficiency by reducing channel-wise redundancy. Skip-scSE applies spatial and channel-wise excitation to signify the essential features and reduces the sparsity in deeper layers learning parameters using skip connections. The proposed approaches are evaluated by comparative experiments with state-of-the-art models in synthesis, segmentation, and overall survival (OS) prediction. We observe that adding missing MRI modality improves the segmentation prediction, and expression levels of gene markers have a high contribution in the GBM prognosis prediction, and fused radiogenomic features boost the OS estimation.