Deep learning techniques have provided significant improvements in hyperspectral image (HSI) classification. The current deep learning based HSI classifiers follow a patch-based learning framework by dividing the image into overlapping patches. As such, these methods are local learning methods, which have a high computational cost. In this paper, a fast patch-free global learning (FPGA) framework is proposed for HSI classification. In FPGA, an encoder-decoder based FCN is utilized to consider the global spatial information by processing the whole image, which results in fast inference. However, it is difficult to directly utilize the encoder-decoder based FCN for HSI classification as it always fails to converge due to the insufficiently diverse gradients caused by the limited training samples. To solve the divergence problem and maintain the abilities of FCN of fast inference and global spatial information mining, a global stochastic stratified sampling strategy is first proposed by transforming all the training samples into a stochastic sequence of stratified samples. This strategy can obtain diverse gradients to guarantee the convergence of the FCN in the FPGA framework. For a better design of FCN architecture, FreeNet, which is a fully end-to-end network for HSI classification, is proposed to maximize the exploitation of the global spatial information and boost the performance via a spectral attention based encoder and a lightweight decoder. A lateral connection module is also designed to connect the encoder and decoder, fusing the spatial details in the encoder and the semantic features in the decoder. The experimental results obtained using three public benchmark datasets suggest that the FPGA framework is superior to the patch-based framework in both speed and accuracy for HSI classification. Code has been made available at: https://github.com/Z-Zheng/FreeNet.
There are limited studies on the semantic segmentation of high-resolution Polarimetric Synthetic Aperture Radar (PolSAR) images due to the scarcity of training data and the inference of speckle noises. The Gaofen contest has provided open access of a high-quality PolSAR semantic segmentation dataset. Taking this chance, we propose a Multi-path ResNet (MP-ResNet) architecture for the semantic segmentation of high-resolution PolSAR images. Compared to conventional U-shape encoder-decoder convolutional neural network (CNN) architectures, the MP-ResNet learns semantic context with its parallel multi-scale branches, which greatly enlarges its valid receptive fields and improves the embedding of local discriminative features. In addition, MP-ResNet adopts a multi-level feature fusion design in its decoder to make the best use of the features learned from its different branches. Ablation studies show that the MPResNet has significant advantages over its baseline method (FCN with ResNet34). It also surpasses several classic state-of-the-art methods in terms of overall accuracy (OA), mean F1 and fwIoU, whereas its computational costs are not much increased. This CNN architecture can be used as a baseline method for future studies on the semantic segmentation of PolSAR images. The code is available at: https://github.com/ggsDing/SARSeg.
Co-Salient Object Detection (CoSOD) aims at discovering salient objects that repeatedly appear in a given query group containing two or more relevant images. One challenging issue is how to effectively capture co-saliency cues by modeling and exploiting inter-image relationships. In this paper, we present an end-to-end collaborative aggregation-and-distribution network (CoADNet) to capture both salient and repetitive visual patterns from multiple images. First, we integrate saliency priors into the backbone features to suppress the redundant background information through an online intra-saliency guidance structure. After that, we design a two-stage aggregate-and-distribute architecture to explore group-wise semantic interactions and produce the co-saliency features. In the first stage, we propose a group-attentional semantic aggregation module that models inter-image relationships to generate the group-wise semantic representations. In the second stage, we propose a gated group distribution module that adaptively distributes the learned group semantics to different individuals in a dynamic gating mechanism. Finally, we develop a group consistency preserving decoder tailored for the CoSOD task, which maintains group constraints during feature decoding to predict more consistent full-resolution co-saliency maps. The proposed CoADNet is evaluated on four prevailing CoSOD benchmark datasets, which demonstrates the remarkable performance improvement over ten state-of-the-art competitors.
In medical imaging, outliers can contain hypo/hyper-intensities, minor deformations, or completely altered anatomy. To detect these irregularities it is helpful to learn the features present in both normal and abnormal images. However this is difficult because of the wide range of possible abnormalities and also the number of ways that normal anatomy can vary naturally. As such, we leverage the natural variations in normal anatomy to create a range of synthetic abnormalities. Specifically, the same patch region is extracted from two independent samples and replaced with an interpolation between both patches. The interpolation factor, patch size, and patch location are randomly sampled from uniform distributions. A wide residual encoder decoder is trained to give a pixel-wise prediction of the patch and its interpolation factor. This encourages the network to learn what features to expect normally and to identify where foreign patterns have been introduced. The estimate of the interpolation factor lends itself nicely to the derivation of an outlier score. Meanwhile the pixel-wise output allows for pixel- and subject- level predictions using the same model.
Automatic segmentation of brain tumors is an essential but challenging step for extracting quantitative imaging biomarkers for accurate tumor detection, diagnosis, prognosis, treatment planning and assessment. Multimodal Brain Tumor Segmentation Challenge 2020 (BraTS 2020) provides a common platform for comparing different automatic algorithms on multi-parametric Magnetic Resonance Imaging (mpMRI) in tasks of 1) Brain tumor segmentation MRI scans; 2) Prediction of patient overall survival (OS) from pre-operative MRI scans; 3) Distinction of true tumor recurrence from treatment related effects and 4) Evaluation of uncertainty measures in segmentation. We participate the image segmentation challenge by developing a fully automatic segmentation network based on encoder-decoder architecture. In order to better integrate information across different scales, we propose a dynamic scale attention mechanism that incorporates low-level details with high-level semantics from feature maps at different scales. Our framework was trained using the 369 challenge training cases provided by BraTS 2020, and achieved an average Dice Similarity Coefficient (DSC) of 0.8828, 0.8433 and 0.8177, as well as 95% Hausdorff distance (in millimeter) of 5.2176, 17.9697 and 13.4298 on 166 testing cases for whole tumor, tumor core and enhanced tumor, respectively, which ranked itself as the 3rd place among 693 registrations in the BraTS 2020 challenge.
Automatic MRI brain tumor segmentation is of vital importance for the disease diagnosis, monitoring, and treatment planning. In this paper, we propose a two-stage encoder-decoder based model for brain tumor subregional segmentation. Variational autoencoder regularization is utilized in both stages to prevent the overfitting issue. The second-stage network adopts attention gates and is trained additionally using an expanded dataset formed by the first-stage outputs. On the BraTS 2020 validation dataset, the proposed method achieves the mean Dice score of 0.9041, 0.8350, and 0.7958, and Hausdorff distance (95%) of 4.953, 6.299, and 23.608 for the whole tumor, tumor core, and enhancing tumor, respectively. The corresponding results on the BraTS 2020 testing dataset are 0.8729, 0.8357, and 0.8205 for Dice score, and 11.4288, 19.9690, and 15.6711 for Hausdorff distance. The code is publicly available at https://github.com/shu-hai/two-stage-VAE-Attention-gate-BraTS2020.
Local ancestry inference (LAI) identifies the ancestry of each segment of an individual's genome and is an important step in medical and population genetic studies of diverse cohorts. Several techniques have been used for LAI, including Hidden Markov Models and Random Forests. Here, we formulate the LAI task as an image segmentation problem and develop a new LAI tool using a deep convolutional neural network with an encoder-decoder architecture. We train our model using complete genome sequences from 982 unadmixed individuals from each of five continental ancestry groups, and we evaluate it using simulated admixed data derived from an additional 279 individuals selected from the same populations. We show that our model is able to learn admixture as a zero-shot task, yielding ancestry assignments that are nearly as accurate as those from the existing gold standard tool, RFMix.
We introduce dual-decoder Transformer, a new model architecture that jointly performs automatic speech recognition (ASR) and multilingual speech translation (ST). Our models are based on the original Transformer architecture (Vaswani et al., 2017) but consist of two decoders, each responsible for one task (ASR or ST). Our major contribution lies in how these decoders interact with each other: one decoder can attend to different information sources from the other via a dual-attention mechanism. We propose two variants of these architectures corresponding to two different levels of dependencies between the decoders, called the parallel and cross dual-decoder Transformers, respectively. Extensive experiments on the MuST-C dataset show that our models outperform the previously-reported highest translation performance in the multilingual settings, and outperform as well bilingual one-to-one results. Furthermore, our parallel models demonstrate no trade-off between ASR and ST compared to the vanilla multi-task architecture. Our code and pre-trained models are available at https://github.com/formiel/speech-translation.
We propose a novel architecture for labelling character sequences that achieves state-of-the-art results on the Tashkeela Arabic diacritization benchmark. The core is a two-level recurrence hierarchy that operates on the word and character levels separately---enabling faster training and inference than comparable traditional models. A cross-level attention module further connects the two, and opens the door for network interpretability. The task module is a softmax classifier that enumerates valid combinations of diacritics. This architecture can be extended with a recurrent decoder that optionally accepts priors from partially diacritized text, which improves results. We employ extra tricks such as sentence dropout and majority voting to further boost the final result. Our best model achieves a WER of 5.34%, outperforming the previous state-of-the-art with a 30.56% relative error reduction.
Object pose estimation enables robots to understand and interact with their environments. Training with synthetic data is necessary in order to adapt to novel situations. Unfortunately, pose estimation under domain shift, i.e., training on synthetic data and testing in the real world, is challenging. Deep learning-based approaches currently perform best when using encoder-decoder networks but typically do not generalize to new scenarios with different scene characteristics. We argue that patch-based approaches, instead of encoder-decoder networks, are more suited for synthetic-to-real transfer because local to global object information is better represented. To that end, we present a novel approach based on a specialized feature pyramid network to compute multi-scale features for creating pose hypotheses on different feature map resolutions in parallel. Our single-shot pose estimation approach is evaluated on multiple standard datasets and outperforms the state of the art by up to 35%. We also perform grasping experiments in the real world to demonstrate the advantage of using synthetic data to generalize to novel environments.
Medical imaging is frequently used in clinical practice and trials for diagnosis and treatment. Writing imaging reports is time-consuming and can be error-prone for inexperienced radiologists. Therefore, automatically generating radiology reports is highly desired to lighten the workload of radiologists and accordingly promote clinical automation, which is an essential task to apply artificial intelligence to the medical domain. In this paper, we propose to generate radiology reports with memory-driven Transformer, where a relational memory is designed to record key information of the generation process and a memory-driven conditional layer normalization is applied to incorporating the memory into the decoder of Transformer. Experimental results on two prevailing radiology report datasets, IU X-Ray and MIMIC-CXR, show that our proposed approach outperforms previous models with respect to both language generation metrics and clinical evaluations. Particularly, this is the first work reporting the generation results on MIMIC-CXR to the best of our knowledge. Further analyses also demonstrate that our approach is able to generate long reports with necessary medical terms as well as meaningful image-text attention mappings.
Question answering is an effective method for obtaining information from knowledge bases (KB). In this paper, we propose the Neural-Symbolic Complex Question Answering (NS-CQA) model, a data-efficient reinforcement learning framework for complex question answering by using only a modest number of training samples. Our framework consists of a neural generator and a symbolic executor that, respectively, transforms a natural-language question into a sequence of primitive actions, and executes them over the knowledge base to compute the answer. We carefully formulate a set of primitive symbolic actions that allows us to not only simplify our neural network design but also accelerate model convergence. To reduce search space, we employ the copy and masking mechanisms in our encoder-decoder architecture to drastically reduce the decoder output vocabulary and improve model generalizability. We equip our model with a memory buffer that stores high-reward promising programs. Besides, we propose an adaptive reward function. By comparing the generated trial with the trials stored in the memory buffer, we derive the curriculum-guided reward bonus, i.e., the proximity and the novelty. To mitigate the sparse reward problem, we combine the adaptive reward and the reward bonus, reshaping the sparse reward into dense feedback. Also, we encourage the model to generate new trials to avoid imitating the spurious trials while making the model remember the past high-reward trials to improve data efficiency. Our NS-CQA model is evaluated on two datasets: CQA, a recent large-scale complex question answering dataset, and WebQuestionsSP, a multi-hop question answering dataset. On both datasets, our model outperforms the state-of-the-art models. Notably, on CQA, NS-CQA performs well on questions with higher complexity, while only using approximately 1% of the total training samples.
Existing object detection frameworks are usually built on a single format of object/part representation, i.e., anchor/proposal rectangle boxes in RetinaNet and Faster R-CNN, center points in FCOS and RepPoints, and corner points in CornerNet. While these different representations usually drive the frameworks to perform well in different aspects, e.g., better classification or finer localization, it is in general difficult to combine these representations in a single framework to make good use of each strength, due to the heterogeneous or non-grid feature extraction by different representations. This paper presents an attention-based decoder module similar as that in Transformer~\cite{vaswani2017attention} to bridge other representations into a typical object detector built on a single representation format, in an end-to-end fashion. The other representations act as a set of \emph{key} instances to strengthen the main \emph{query} representation features in the vanilla detectors. Novel techniques are proposed towards efficient computation of the decoder module, including a \emph{key sampling} approach and a \emph{shared location embedding} approach. The proposed module is named \emph{bridging visual representations} (BVR). It can perform in-place and we demonstrate its broad effectiveness in bridging other representations into prevalent object detection frameworks, including RetinaNet, Faster R-CNN, FCOS and ATSS, where about $1.5\sim3.0$ AP improvements are achieved. In particular, we improve a state-of-the-art framework with a strong backbone by about $2.0$ AP, reaching $52.7$ AP on COCO test-dev. The resulting network is named RelationNet++. The code will be available at https://github.com/microsoft/RelationNet2.
Emotional voice conversion aims to transform emotional prosody in speech while preserving the linguistic content and speaker identity. Prior studies show that it is possible to disentangle emotional prosody using an encoder-decoder network conditioned on discrete representation, such as one-hot emotion labels. Such networks learn to remember a fixed set of emotional styles. In this paper, we propose a novel framework based on variational auto-encoding Wasserstein generative adversarial network (VAW-GAN), which makes use of a pre-trained speech emotion recognition (SER) model to transfer emotional style during training and at run-time inference. In this way, the network is able to transfer both seen and unseen emotional style to a new utterance. We show that the proposed framework achieves remarkable performance by consistently outperforming the baseline framework. This paper also marks the release of an emotional speech dataset (ESD) for voice conversion, which has multiple speakers and languages.
We propose a new approach for learning contextualised cross-lingual word embeddings based on a small parallel corpus (e.g. a few hundred sentence pairs). Our method obtains word embeddings via an LSTM encoder-decoder model that simultaneously translates and reconstructs an input sentence. Through sharing model parameters among different languages, our model jointly trains the word embeddings in a common cross-lingual space. We also propose to combine word and subword embeddings to make use of orthographic similarities across different languages. We base our experiments on real-world data from endangered languages, namely Yongning Na, Shipibo-Konibo, and Griko. Our experiments on bilingual lexicon induction and word alignment tasks show that our model outperforms existing methods by a large margin for most language pairs. These results demonstrate that, contrary to common belief, an encoder-decoder translation model is beneficial for learning cross-lingual representations even in extremely low-resource conditions. Furthermore, our model also works well on high-resource conditions, achieving state-of-the-art performance on a German-English word-alignment task.
We investigate the problem of multi-domain Dialogue State Tracking (DST) with open vocabulary. Existing approaches exploit BERT encoder and copy-based RNN decoder, where the encoder predicts the state operation, and the decoder generates new slot values. However, in such a stacked encoder-decoder structure, the operation prediction objective only affects the BERT encoder and the value generation objective mainly affects the RNN decoder. In this paper, we propose a purely Transformer-based framework, where a single BERT works as both the encoder and the decoder. In so doing, the operation prediction objective and the value generation objective can jointly optimize this BERT for DST. At the decoding step, we re-use the hidden states of the encoder in the self-attention mechanism of the corresponding decoder layers to construct a flat encoder-decoder architecture for effective parameter updating. Experimental results show that our approach substantially outperforms the existing state-of-the-art framework, and it also achieves very competitive performance to the best ontology-based approaches.
We consider a new perspective on dialog state tracking (DST), the task of estimating a user's goal through the course of a dialog. By formulating DST as a semantic parsing task over hierarchical representations, we can incorporate semantic compositionality, cross-domain knowledge sharing and co-reference. We present TreeDST, a dataset of 27k conversations annotated with tree-structured dialog states and system acts. We describe an encoder-decoder framework for DST with hierarchical representations, which leads to 20% improvement over state-of-the-art DST approaches that operate on a flat meaning space of slot-value pairs.
Road surface extraction from remote sensing images using deep learning methods has achieved good performance, while most of the existing methods are based on fully supervised learning, which requires a large amount of training data with laborious per-pixel annotation. In this paper, we propose a scribble-based weakly supervised road surface extraction method named ScRoadExtractor, which learns from easily accessible scribbles such as centerlines instead of densely annotated road surface ground-truths. To propagate semantic information from sparse scribbles to unlabeled pixels, we introduce a road label propagation algorithm which considers both the buffer-based properties of road networks and the color and spatial information of super-pixels. The proposal masks generated from the road label propagation algorithm are utilized to train a dual-branch encoder-decoder network we designed, which consists of a semantic segmentation branch and an auxiliary boundary detection branch. We perform experiments on three diverse road datasets that are comprised of highresolution remote sensing satellite and aerial images across the world. The results demonstrate that ScRoadExtractor exceed the classic scribble-supervised segmentation method by 20% for the intersection over union (IoU) indicator and outperform the state-of-the-art scribble-based weakly supervised methods at least 4%.
Many neural network models nowadays have achieved promising performances in Chit-chat settings. The majority of them rely on an encoder for understanding the post and a decoder for generating the response. Without given assigned semantics, the models lack the fine-grained control over responses as the semantic mapping between posts and responses is hidden on the fly within the end-to-end manners. Some previous works utilize sampled latent words as a controllable semantic form to drive the generated response around the work, but few works attempt to use more complex semantic patterns to guide the generation. In this paper, we propose to use more detailed semantic forms, including latent responses and part-of-speech sequences sampled from the corresponding distributions, as the controllable semantics to guide the generation. Our results show that the richer semantics are not only able to provide informative and diverse responses, but also increase the overall performance of response quality, including fluency and coherence.
Independence assumptions during sequence generation can speed up inference, but parallel generation of highly inter-dependent tokens comes at a cost in quality. Instead of assuming independence between neighbouring tokens (semi-autoregressive decoding, SA), we take inspiration from bidirectional sequence generation and introduce a decoder that generates target words from the left-to-right and right-to-left directions simultaneously. We show that we can easily convert a standard architecture for unidirectional decoding into a bidirectional decoder by simply interleaving the two directions and adapting the word positions and self-attention masks. Our interleaved bidirectional decoder (IBDecoder) retains the model simplicity and training efficiency of the standard Transformer, and on five machine translation tasks and two document summarization tasks, achieves a decoding speedup of ~2X compared to autoregressive decoding with comparable quality. Notably, it outperforms left-to-right SA because the independence assumptions in IBDecoder are more felicitous. To achieve even higher speedups, we explore hybrid models where we either simultaneously predict multiple neighbouring tokens per direction, or perform multi-directional decoding by partitioning the target sequence. These methods achieve speedups to 4X-11X across different tasks at the cost of <1 BLEU or <0.5 ROUGE (on average). Source code is released at https://github.com/bzhangGo/zero.