The celebrated Seq2Seq technique and its numerous variants achieve excellent performance on many tasks such as neural machine translation, semantic parsing, and math word problem solving. However, these models either only consider input objects as sequences while ignoring the important structural information for encoding, or they simply treat output objects as sequence outputs instead of structural objects for decoding. In this paper, we present a novel Graph-to-Tree Neural Networks, namely Graph2Tree consisting of a graph encoder and a hierarchical tree decoder, that encodes an augmented graph-structured input and decodes a tree-structured output. In particular, we investigated our model for solving two problems, neural semantic parsing and math word problem. Our extensive experiments demonstrate that our Graph2Tree model outperforms or matches the performance of other state-of-the-art models on these tasks.
Residual network (ResNet) and densely connected network (DenseNet) have significantly improved the training efficiency and performance of deep convolutional neural networks (DCNNs) mainly for object classification tasks. In this paper, we propose an efficient network architecture by considering advantages of both networks. The proposed method is integrated into an encoder-decoder DCNN model for medical image segmentation. Our method adds additional skip connections compared to ResNet but uses significantly fewer model parameters than DenseNet. We evaluate the proposed method on a public dataset (ISIC 2018 grand-challenge) for skin lesion segmentation and a local brain MRI dataset. In comparison with ResNet-based, DenseNet-based and attention network (AttnNet) based methods within the same encoder-decoder network structure, our method achieves significantly higher segmentation accuracy with fewer number of model parameters than DenseNet and AttnNet. The code is available on GitHub (GitHub link: https://github.com/MinaJf/DRU-net).
In this paper, we propose a Multi-Scale Boosted Dehazing Network with Dense Feature Fusion based on the U-Net architecture. The proposed method is designed based on two principles, boosting and error feedback, and we show that they are suitable for the dehazing problem. By incorporating the Strengthen-Operate-Subtract boosting strategy in the decoder of the proposed model, we develop a simple yet effective boosted decoder to progressively restore the haze-free image. To address the issue of preserving spatial information in the U-Net architecture, we design a dense feature fusion module using the back-projection feedback scheme. We show that the dense feature fusion module can simultaneously remedy the missing spatial information from high-resolution features and exploit the non-adjacent features. Extensive evaluations demonstrate that the proposed model performs favorably against the state-of-the-art approaches on the benchmark datasets as well as real-world hazy images.
In neural network-based models for natural language processing (NLP), the largest part of the parameters often consists of word embeddings. Conventional models prepare a large embedding matrix whose size depends on the vocabulary size. Therefore, storing these models in memory and disk storage is costly. In this study, to reduce the total number of parameters, the embeddings for all words are represented by transforming a shared embedding. The proposed method, ALONE (all word embeddings from one), constructs the embedding of a word by modifying the shared embedding with a filter vector, which is word-specific but non-trainable. Then, we input the constructed embedding into a feed-forward neural network to increase its expressiveness. Naively, the filter vectors occupy the same memory size as the conventional embedding matrix, which depends on the vocabulary size. To solve this issue, we also introduce a memory-efficient filter construction approach. We indicate our ALONE can be used as word representation sufficiently through an experiment on the reconstruction of pre-trained word embeddings. In addition, we also conduct experiments on NLP application tasks: machine translation and summarization. We combined ALONE with the current state-of-the-art encoder-decoder model, the Transformer, and achieved comparable scores on WMT 2014 English-to-German translation and DUC 2004 very short summarization with less parameters.
Designing effective neural networks is fundamentally important in deep multimodal learning. Most existing works focus on a single task and design neural architectures manually, which are highly task-specific and hard to generalize to different tasks. In this paper, we devise a generalized deep multimodal neural architecture search (MMnas) framework for various multimodal learning tasks. Given multimodal input, we first define a set of primitive operations, and then construct a deep encoder-decoder based unified backbone, where each encoder or decoder block corresponds to an operation searched from a predefined operation pool. On top of the unified backbone, we attach task-specific heads to tackle different multimodal learning tasks. By using a gradient-based NAS algorithm, the optimal architectures for different tasks are learned efficiently. Extensive ablation studies, comprehensive analysis, and comparative experimental results show that the obtained MMnasNet significantly outperforms existing state-of-the-art approaches across three multimodal learning tasks (over five datasets), including visual question answering, image-text matching, and visual grounding.
Change-point detection (CPD) aims to detect abrupt changes over time series data. Intuitively, effective CPD over multivariate time series should require explicit modeling of the dependencies across input variables. However, existing CPD methods either ignore the dependency structures entirely or rely on the (unrealistic) assumption that the correlation structures are static over time. In this paper, we propose a Correlation-aware Dynamics Model for CPD, which explicitly models the correlation structure and dynamics of variables by incorporating graph neural networks into an encoder-decoder framework. Extensive experiments on synthetic and real-world datasets demonstrate the advantageous performance of the proposed model on CPD tasks over strong baselines, as well as its ability to classify the change-points as correlation changes or independent changes. Keywords: Multivariate Time Series, Change-point Detection, Graph Neural Networks
Video Object Segmentation (VOS) is an active research area of the visual domain. One of its fundamental sub-tasks is semi-supervised / one-shot learning: given only the segmentation mask for the first frame, the task is to provide pixel-accurate masks for the object over the rest of the sequence. Despite much progress in the last years, we noticed that many of the existing approaches lose objects in longer sequences, especially when the object is small or briefly occluded. In this work, we build upon a sequence-to-sequence approach that employs an encoder-decoder architecture together with a memory module for exploiting the sequential data. We further improve this approach by proposing a model that manipulates multi-scale spatio-temporal information using memory-equipped skip connections. Furthermore, we incorporate an auxiliary task based on distance classification which greatly enhances the quality of edges in segmentation masks. We compare our approach to the state of the art and show considerable improvement in the contour accuracy metric and the overall segmentation accuracy.
Benefited from the deep learning, image Super-Resolution has been one of the most developing research fields in computer vision. Depending upon whether using a discriminator or not, a deep convolutional neural network can provide an image with high fidelity or better perceptual quality. Due to the lack of ground truth images in real life, people prefer a photo-realistic image with low fidelity to a blurry image with high fidelity. In this paper, we revisit the classic example based image super-resolution approaches and come up with a novel generative model for perceptual image super-resolution. Given that real images contain various noise and artifacts, we propose a joint image denoising and super-resolution model via Variational AutoEncoder. We come up with a conditional variational autoencoder to encode the reference for dense feature vector which can then be transferred to the decoder for target image denoising. With the aid of the discriminator, an additional overhead of super-resolution subnetwork is attached to super-resolve the denoised image with photo-realistic visual quality. We participated the NTIRE2020 Real Image Super-Resolution Challenge. Experimental results show that by using the proposed approach, we can obtain enlarged images with clean and pleasant features compared to other supervised methods. We also compared our approach with state-of-the-art methods on various datasets to demonstrate the efficiency of our proposed unsupervised super-resolution model.
Advances in deep learning techniques have allowed recent work to reconstruct the shape of a single object given only one RBG image as input. Building on common encoder-decoder architectures for this task, we propose three extensions: (1) ray-traced skip connections that propagate local 2D information to the output 3D volume in a physically correct manner; (2) a hybrid 3D volume representation that enables building translation equivariant models, while at the same time encoding fine object details without an excessive memory footprint; (3) a reconstruction loss tailored to capture overall object geometry. Furthermore, we adapt our model to address the harder task of reconstructing multiple objects from a single image. We reconstruct all objects jointly in one pass, producing a coherent reconstruction, where all objects live in a single consistent 3D coordinate frame relative to the camera and they do not intersect in 3D space. We also handle occlusions and resolve them by hallucinating the missing object parts in the 3D volume. We validate the impact of our contributions experimentally both on synthetic data from ShapeNet as well as real images from Pix3D. Our method improves over the state-of-the-art single-object methods on both datasets. Finally, we evaluate performance quantitatively on multiple object reconstruction with synthetic scenes assembled from ShapeNet objects.
In this work, we propose a new generative model that is capable of automatically decoupling global and local representations of images in an entirely unsupervised setting, by embedding a generative flow in the VAE framework to model the decoder. Specifically, the proposed model utilizes the variational auto-encoding framework to learn a (low-dimensional) vector of latent variables to capture the global information of an image, which is fed as a conditional input to a flow-based invertible decoder with architecture borrowed from style transfer literature. Experimental results on standard image benchmarks demonstrate the effectiveness of our model in terms of density estimation, image generation and unsupervised representation learning. Importantly, this work demonstrates that with only architectural inductive biases, a generative model with a likelihood-based objective is capable of learning decoupled representations, requiring no explicit supervision. The code for our model is available at https://github.com/XuezheMax/wolf.
Sequence-to-sequence models usually transfer all encoder outputs to the decoder for generation. In this work, by contrast, we hypothesize that these encoder outputs can be compressed to shorten the sequence delivered for decoding. We take Transformer as the testbed and introduce a layer of stochastic gates in-between the encoder and the decoder. The gates are regularized using the expected value of the sparsity-inducing L0penalty, resulting in completely masking-out a subset of encoder outputs. In other words, via joint training, the L0DROP layer forces Transformer to route information through a subset of its encoder states. We investigate the effects of this sparsification on two machine translation and two summarization tasks. Experiments show that, depending on the task, around 40-70% of source encodings can be pruned without significantly compromising quality. The decrease of the output length endows L0DROP with the potential of improving decoding efficiency, where it yields a speedup of up to 1.65x on document summarization tasks against the standard Transformer. We analyze the L0DROP behaviour and observe that it exhibits systematic preferences for pruning certain word types, e.g., function words and punctuation get pruned most. Inspired by these observations, we explore the feasibility of specifying rule-based patterns that mask out encoder outputs based on information such as part-of-speech tags, word frequency and word position.
Keyphrase generation aims to produce a set of phrases summarizing the essentials of a given document. Conventional methods normally apply an encoder-decoder architecture to generate the output keyphrases for an input document, where they are designed to focus on each current document so they inevitably omit crucial corpus-level information carried by other similar documents, i.e., the cross-document dependency and latent topics. In this paper, we propose CDKGen, a Transformer-based keyphrase generator, which expands the Transformer to global attention with cross-document attention networks to incorporate available documents as references so as to generate better keyphrases with the guidance of topic information. On top of the proposed Transformer + cross-document attention architecture, we also adopt a copy mechanism to enhance our model via selecting appropriate words from documents to deal with out-of-vocabulary words in keyphrases. Experiment results on five benchmark datasets illustrate the validity and effectiveness of our model, which achieves the state-of-the-art performance on all datasets. Further analyses confirm that the proposed model is able to generate keyphrases consistent with references while keeping sufficient diversity. The code of CDKGen is available at https://github.com/SVAIGBA/CDKGen.
Variational autoencoders (VAEs) combine latent variables with amortized variational inference, whose optimization usually converges into a trivial local optimum termed posterior collapse, especially in text modeling. By tracking the optimization dynamics, we observe the encoder-decoder incompatibility that leads to poor parameterizations of the data manifold. We argue that the trivial local optimum may be avoided by improving the encoder and decoder parameterizations since the posterior network is part of a transition map between them. To this end, we propose Coupled-VAE, which couples a VAE model with a deterministic autoencoder with the same structure and improves the encoder and decoder parameterizations via encoder weight sharing and decoder signal matching. We apply the proposed Coupled-VAE approach to various VAE models with different regularization, posterior family, decoder structure, and optimization strategy. Experiments on benchmark datasets (i.e., PTB, Yelp, and Yahoo) show consistently improved results in terms of probability estimation and richness of the latent space. We also generalize our method to conditional language modeling and propose Coupled-CVAE, which largely improves the diversity of dialogue generation on the Switchboard dataset.
Recently, a growing interest has been seen in deep learning-based semantic segmentation. UNet, which is one of deep learning networks with an encoder-decoder architecture, is widely used in medical image segmentation. Combining multi-scale features is one of important factors for accurate segmentation. UNet++ was developed as a modified Unet by designing an architecture with nested and dense skip connections. However, it does not explore sufficient information from full scales and there is still a large room for improvement. In this paper, we propose a novel UNet 3+, which takes advantage of full-scale skip connections and deep supervisions. The full-scale skip connections incorporate low-level details with high-level semantics from feature maps in different scales; while the deep supervision learns hierarchical representations from the full-scale aggregated feature maps. The proposed method is especially benefiting for organs that appear at varying scales. In addition to accuracy improvements, the proposed UNet 3+ can reduce the network parameters to improve the computation efficiency. We further propose a hybrid loss function and devise a classification-guided module to enhance the organ boundary and reduce the over-segmentation in a non-organ image, yielding more accurate segmentation results. The effectiveness of the proposed method is demonstrated on two datasets. The code is available at: github.com/ZJUGiveLab/UNet-Version
Deep Neural Networks (DNN) have been widely used to carry out segmentation tasks in both electron and light microscopy. Most DNNs developed for this purpose are based on some variation of the encoder-decoder type U-Net architecture, in combination with residual blocks to increase ease of training and resilience to gradient degradation. Here we introduce Res-CR-Net, a type of DNN that features residual blocks with either a bundle of separable atrous convolutions with different dilation rates or a convolutional LSTM. The number of filters used in each residual block and the number of blocks are the only hyperparameters that need to be modified in order to optimize the network training for a variety of different microscopy images.
Speaker extraction aims to mimic humans' selective auditory attention by extracting a target speaker's voice from a multi-talker environment. It is common to perform the extraction in frequency-domain, and reconstruct the time-domain signal from the extracted magnitude and estimated phase spectra. However, such an approach is adversely affected by the inherent difficulty of phase estimation. Inspired by Conv-TasNet, we propose a time-domain speaker extraction network (SpEx) that converts the mixture speech into multi-scale embedding coefficients instead of decomposing the speech signal into magnitude and phase spectra. In this way, we avoid phase estimation. The SpEx network consists of four network components, namely speaker encoder, speech encoder, speaker extractor, and speech decoder. Specifically, the speech encoder converts the mixture speech into multi-scale embedding coefficients, the speaker encoder learns to represent the target speaker with a speaker embedding. The speaker extractor takes the multi-scale embedding coefficients and target speaker embedding as input and estimates a receptive mask. Finally, the speech decoder reconstructs the target speaker's speech from the masked embedding coefficients. We also propose a multi-task learning framework and a multi-scale embedding implementation. Experimental results show that the proposed SpEx achieves 37.3%, 37.7% and 15.0% relative improvements over the best baseline in terms of signal-to-distortion ratio (SDR), scale-invariant SDR (SI-SDR), and perceptual evaluation of speech quality (PESQ) under an open evaluation condition.
Deep learning based image segmentation has achieved the state-of-the-art performance in many medical applications such as lesion quantification, organ detection, etc. However, most of the methods rely on supervised learning, which require a large set of high-quality labeled data. Data annotation is generally an extremely time-consuming process. To address this problem, we propose a generic semi-supervised learning framework for image segmentation based on a deep convolutional neural network (DCNN). An encoder-decoder based DCNN is initially trained using a few annotated training samples. This initially trained model is then copied into sub-models and improved iteratively using random subsets of unlabeled data with pseudo labels generated from models trained in the previous iteration. The number of sub-models is gradually decreased to one in the final iteration. We evaluate the proposed method on a public grand-challenge dataset for skin lesion segmentation. Our method is able to significantly improve beyond fully supervised model learning by incorporating unlabeled data.
The extraction of a scene graph with objects as nodes and mutual relationships as edges is the basis for a deep understanding of image content. Despite recent advances, such as message passing and joint classification, the detection of visual relationships remains a challenging task due to sub-optimal exploration of the mutual interaction among the visual objects. In this work, we propose a novel transformer formulation for scene graph generation and relation prediction. We leverage the encoder-decoder architecture of the transformer for rich feature embedding of nodes and edges. Specifically, we model the node-to-node interaction with the self-attention of the transformer encoder and the edge-to-node interaction with the cross-attention of the transformer decoder. Further, we introduce a novel positional embedding suitable to handle edges in the decoder. Finally, our relation prediction module classifies the directed relation from the learned node and edge embedding. We name this architecture as Relation Transformer Network (RTN). On the Visual Genome and GQA dataset, we have achieved an overall mean of 4.85% and 3.1% point improvement in comparison with state-of-the-art methods. Our experiments show that Relation Transformer can efficiently model context across various datasets with small, medium, and large-scale relation classification.
Transformer-based models are unable to process long sequences due to their self-attention operation, which scales quadratically with the sequence length. To address this limitation, we introduce the Longformer with an attention mechanism that scales linearly with sequence length, making it easy to process documents of thousands of tokens or longer. Longformer's attention mechanism is a drop-in replacement for the standard self-attention and combines a local windowed attention with a task motivated global attention. Following prior work on long-sequence transformers, we evaluate Longformer on character-level language modeling and achieve state-of-the-art results on text8 and enwik8. In contrast to most prior work, we also pretrain Longformer and finetune it on a variety of downstream tasks. Our pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new state-of-the-art results on WikiHop and TriviaQA. We finally introduce the Longformer-Encoder-Decoder (LED), a Longformer variant for supporting long document generative sequence-to-sequence tasks, and demonstrate its effectiveness on the arXiv summarization dataset.
Large pre-trained language models (LMs) are known to encode substantial amounts of linguistic information. However, high-level reasoning skills, such as numerical reasoning, are difficult to learn from a language-modeling objective only. Consequently, existing models for numerical reasoning have used specialized architectures with limited flexibility. In this work, we show that numerical reasoning is amenable to automatic data generation, and thus one can inject this skill into pre-trained LMs, by generating large amounts of data, and training in a multi-task setup. We show that pre-training our model, GenBERT, on this data, dramatically improves performance on DROP (49.3 $\rightarrow$ 72.3 F1), reaching performance that matches state-of-the-art models of comparable size, while using a simple and general-purpose encoder-decoder architecture. Moreover, GenBERT generalizes well to math word problem datasets, while maintaining high performance on standard RC tasks. Our approach provides a general recipe for injecting skills into large pre-trained LMs, whenever the skill is amenable to automatic data augmentation.