Advanced neural machine translation (NMT) models generally implement encoder and decoder as multiple layers, which allows systems to model complex functions and capture complicated linguistic structures. However, only the top layers of encoder and decoder are leveraged in the subsequent process, which misses the opportunity to exploit the useful information embedded in other layers. In this work, we propose to simultaneously expose all of these signals with layer aggregation and multi-layer attention mechanisms. In addition, we introduce an auxiliary regularization term to encourage different layers to capture diverse information. Experimental results on widely-used WMT14 English-German and WMT17 Chinese-English translation data demonstrate the effectiveness and universality of the proposed approach.
Mammography is one of the most commonly applied tools for early breast cancer screening. Automatic segmentation of breast masses in mammograms is essential but challenging due to the low signal-to-noise ratio and the wide variety of mass shapes and sizes. Existing methods deal with these challenges mainly by extracting mass-centered image patches manually or automatically. However, manual patch extraction is time-consuming and automatic patch extraction brings errors that could not be compensated in the following segmentation step. In this study, we propose a novel attention-guided dense-upsampling network (AUNet) for accurate breast mass segmentation in whole mammograms directly. In AUNet, we employ an asymmetrical encoder-decoder structure and propose an effective upsampling block, attention-guided dense-upsampling block (AU block). Especially, the AU block is designed to have three merits. Firstly, it compensates the information loss of bilinear upsampling by dense upsampling. Secondly, it designs a more effective method to fuse high- and low-level features. Thirdly, it includes a channel-attention function to highlight rich-information channels. We evaluated the proposed method on two publicly available datasets, CBIS-DDSM and INbreast. Compared to three state-of-the-art fully convolutional networks, AUNet achieved the best performances with an average Dice similarity coefficient of 81.8% for CBIS-DDSM and 79.1% for INbreast.
Recent progress in deep learning for audio synthesis opens the way to models that directly produce the waveform, shifting away from the traditional paradigm of relying on vocoders or MIDI synthesizers for speech or music generation. Despite their successes, current state-of-the-art neural audio synthesizers such as WaveNet and SampleRNN suffer from prohibitive training and inference times because they are based on autoregressive models that generate audio samples one at a time at a rate of 16kHz. In this work, we study the more computationally efficient alternative of generating the waveform frame-by-frame with large strides. We present SING, a lightweight neural audio synthesizer for the original task of generating musical notes given desired instrument, pitch and velocity. Our model is trained end-to-end to generate notes from nearly 1000 instruments with a single decoder, thanks to a new loss function that minimizes the distances between the log spectrograms of the generated and target waveforms. On the generalization task of synthesizing notes for pairs of pitch and instrument not seen during training, SING produces audio with significantly improved perceptual quality compared to a state-of-the-art autoencoder based on WaveNet as measured by a Mean Opinion Score (MOS), and is about 32 times faster for training and 2, 500 times faster for inference.
In the last years, automated segmentation has become a necessary tool for volume electron microscopy (EM) imaging. So far, the best performing techniques have been largely based on fully supervised encoder-decoder CNNs, requiring a substantial amount of annotated images. Domain Adaptation (DA) aims to alleviate the annotation burden by 'adapting' the networks trained on existing groundtruth data (source domain) to work on a different (target) domain with as little additional annotation as possible. Most DA research is focused on the classification task, whereas volume EM segmentation remains rather unexplored. In this work, we extend recently proposed classification DA techniques to an encoder-decoder layout and propose a novel method that adds a reconstruction decoder to the classical encoder-decoder segmentation in order to align source and target encoder features. The method has been validated on the task of segmenting mitochondria in EM volumes. We have performed DA from brain EM images to HeLa cells and from isotropic FIB/SEM volumes to anisotropic TEM volumes. In all cases, the proposed method has outperformed the extended classification DA techniques and the finetuning baseline. An implementation of our work can be found on https://github.com/JorisRoels/domain-adaptive-segmentation.
Program synthesis from natural language (NL) is practical for humans and, once technically feasible, would significantly facilitate software development and revolutionize end-user programming. We present SAPS, an end-to-end neural network capable of mapping relatively complex, multi-sentence NL specifications to snippets of executable code. The proposed architecture relies exclusively on neural components, and is trained on abstract syntax trees, combined with a pretrained word embedding and a bi-directional multi-layer LSTM for processing of word sequences. The decoder features a doubly-recurrent LSTM, for which we propose novel signal propagation schemes and soft attention mechanism. When applied to a large dataset of problems proposed in a previous study, SAPS performs on par with or better than the method proposed there, producing correct programs in over 92% of cases. In contrast to other methods, it does not require post-processing of the resulting programs, and uses a fixed-dimensional latent representation as the only interface between the NL analyzer and the source code generator.
In a world of proliferating data, the ability to rapidly summarize text is growing in importance. Automatic summarization of text can be thought of as a sequence to sequence problem. Another area of natural language processing that solves a sequence to sequence problem is machine translation, which is rapidly evolving due to the development of attention-based encoder-decoder networks. This work applies these modern techniques to abstractive summarization. We perform analysis on various attention mechanisms for summarization with the goal of developing an approach and architecture aimed at improving the state of the art. In particular, we modify and optimize a translation model with self-attention for generating abstractive sentence summaries. The effectiveness of this base model along with attention variants is compared and analyzed in the context of standardized evaluation sets and test metrics. However, we show that these metrics are limited in their ability to effectively score abstractive summaries, and propose a new approach based on the intuition that an abstractive model requires an abstractive evaluation.
Duplicate removal is a critical step to accomplish a reasonable amount of predictions in prevalent proposal-based object detection frameworks. Albeit simple and effective, most previous algorithms utilize a greedy process without making sufficient use of properties of input data. In this work, we design a new two-stage framework to effectively select the appropriate proposal candidate for each object. The first stage suppresses most of easy negative object proposals, while the second stage selects true positives in the reduced proposal set. These two stages share the same network structure, \ie, an encoder and a decoder formed as recurrent neural networks (RNN) with global attention and context gate. The encoder scans proposal candidates in a sequential manner to capture the global context information, which is then fed to the decoder to extract optimal proposals. In our extensive experiments, the proposed method outperforms other alternatives by a large margin.
U-Net has been providing state-of-the-art performance in many medical image segmentation problems. Many modifications have been proposed for U-Net, such as attention U-Net, recurrent residual convolutional U-Net (R2-UNet), and U-Net with residual blocks or blocks with dense connections. However, all these modifications have an encoder-decoder structure with skip connections, and the number of paths for information flow is limited. We propose LadderNet in this paper, which can be viewed as a chain of multiple U-Nets. Instead of only one pair of encoder branch and decoder branch in U-Net, a LadderNet has multiple pairs of encoder-decoder branches, and has skip connections between every pair of adjacent decoder and decoder branches in each level. Inspired by the success of ResNet and R2-UNet, we use modified residual blocks where two convolutional layers in one block share the same weights. A LadderNet has more paths for information flow because of skip connections and residual blocks, and can be viewed as an ensemble of Fully Convolutional Networks (FCN). The equivalence to an ensemble of FCNs improves segmentation accuracy, while the shared weights within each residual block reduce parameter number. Semantic segmentation is essential for retinal disease detection. We tested LadderNet on two benchmark datasets for blood vessel segmentation in retinal images, and achieved superior performance over methods in the literature. The implementation is provided \url{https://github.com/juntang-zhuang/LadderNet}
We suggest a new NLG task in the context of the discourse generation pipeline of computational storytelling systems. This task, textual embellishment, is defined by taking a text as input and generating a semantically equivalent output with increased lexical and syntactic complexity. Ideally, this would allow the authors of computational storytellers to implement just lightweight NLG systems and use a domain-independent embellishment module to translate its output into more literary text. We present promising first results on this task using LSTM Encoder-Decoder networks trained on the WikiLarge dataset. Furthermore, we introduce "Compiled Computer Tales", a corpus of computationally generated stories, that can be used to test the capabilities of embellishment algorithms.
Recent work has shown that the encoder-decoder attention mechanisms in neural machine translation (NMT) are different from the word alignment in statistical machine translation. In this paper, we focus on analyzing encoder-decoder attention mechanisms, in the case of word sense disambiguation (WSD) in NMT models. We hypothesize that attention mechanisms pay more attention to context tokens when translating ambiguous words. We explore the attention distribution patterns when translating ambiguous nouns. Counter-intuitively, we find that attention mechanisms are likely to distribute more attention to the ambiguous noun itself rather than context tokens, in comparison to other nouns. We conclude that attention mechanism is not the main mechanism used by NMT models to incorporate contextual information for WSD. The experimental results suggest that NMT models learn to encode contextual information necessary for WSD in the encoder hidden states. For the attention mechanism in Transformer models, we reveal that the first few layers gradually learn to "align" source and target tokens and the last few layers learn to extract features from the related but unaligned context tokens.
Idioms pose problems to almost all Machine Translation systems. This type of language is very frequent in day-to-day language use and cannot be simply ignored. The recent interest in memory augmented models in the field of Language Modelling has aided the systems to achieve good results by bridging long-distance dependencies. In this paper we explore the use of such techniques into a Neural Machine Translation system to help in translation of idiomatic language.
In this paper, we estimate perceived image quality using sparse representations obtained from generic image databases through an unsupervised learning approach. A color space transformation, a mean subtraction, and a whitening operation are used to enhance descriptiveness of images by reducing spatial redundancy; a linear decoder is used to obtain sparse representations; and a thresholding stage is used to formulate suppression mechanisms in a visual system. A linear decoder is trained with 7 GB worth of data, which corresponds to 100,000 8x8 image patches randomly obtained from nearly 1,000 images in the ImageNet 2013 database. A patch-wise training approach is preferred to maintain local information. The proposed quality estimator UNIQUE is tested on the LIVE, the Multiply Distorted LIVE, and the TID 2013 databases and compared with thirteen quality estimators. Experimental results show that UNIQUE is generally a top performing quality estimator in terms of accuracy, consistency, linearity, and monotonic behavior.
Pronouns are frequently omitted in pro-drop languages, such as Chinese, generally leading to significant challenges with respect to the production of complete translations. Recently, Wang et al. (2018) proposed a novel reconstruction-based approach to alleviating dropped pronoun (DP) translation problems for neural machine translation models. In this work, we improve the original model from two perspectives. First, we employ a shared reconstructor to better exploit encoder and decoder representations. Second, we jointly learn to translate and predict DPs in an end-to-end manner, to avoid the errors propagated from an external DP prediction model. Experimental results show that our approach significantly improves both translation performance and DP prediction accuracy.
We present a novel, parameter-efficient and practical fully convolutional neural network architecture, termed InfiNet, aimed at voxel-wise semantic segmentation of infant brain MRI images at iso-intense stage, which can be easily extended for other segmentation tasks involving multi-modalities. InfiNet consists of double encoder arms for T1 and T2 input scans that feed into a joint-decoder arm that terminates in the classification layer. The novelty of InfiNet lies in the manner in which the decoder upsamples lower resolution input feature map(s) from multiple encoder arms. Specifically, the pooled indices computed in the max-pooling layers of each of the encoder blocks are related to the corresponding decoder block to perform non-linear learning-free upsampling. The sparse maps are concatenated with intermediate encoder representations (skip connections) and convolved with trainable filters to produce dense feature maps. InfiNet is trained end-to-end to optimize for the Generalized Dice Loss, which is well-suited for high class imbalance. InfiNet achieves the whole-volume segmentation in under 50 seconds and we demonstrate competitive performance against multiple state-of-the art deep architectures and their multi-modal variants.
Deep learning models have shown state-of-the-art performance in many inverse reconstruction problems. However, it is not well understood what properties of the latent representation may improve the generalization ability of the network. Furthermore, limited models have been presented for inverse reconstructions over time sequences. In this paper, we study the generalization ability of a sequence encoder decoder model for solving inverse reconstructions on time sequences. Our central hypothesis is that the generalization ability of the network can be improved by 1) constrained stochasticity and 2) global aggregation of temporal information in the latent space. First, drawing from analytical learning theory, we theoretically show that a stochastic latent space will lead to an improved generalization ability. Second, we consider an LSTM encoder-decoder architecture that compresses a global latent vector from all last-layer units in the LSTM encoder. This model is compared with alternative LSTM encoder-decoder architectures, each in deterministic and stochastic versions. The results demonstrate that the generalization ability of an inverse reconstruction network can be improved by constrained stochasticity combined with global aggregation of temporal information in the latent space.
The advent of deep perceptual networks brought about a paradigm shift in machine vision and image perception. Image apprehension lately carried out by hand-crafted features in the latent space have been replaced by deep features acquired from supervised networks for improved understanding. However, such deep networks require strict supervision with a substantial amount of the labeled data for authentic training process. These methods perform poorly in domains lacking labeled data especially in case of remote sensing image retrieval. Resolving this, we propose an unsupervised encoder-decoder feature for remote sensing image matching (RSIM). Moreover, we replace the conventional distance metrics with a deep discriminator network to identify the similarity of the image pairs. To the best of our knowledge, discriminator network has never been used before for solving RSIM problem. Results have been validated with two publicly available benchmark remote sensing image datasets. The technique has also been investigated for content-based remote sensing image retrieval (CBRSIR); one of the widely used applications of RSIM. Results demonstrate that our technique supersedes the state-of-the-art methods used for unsupervised image matching with mean average precision (mAP) of 81%, and image retrieval with an overall improvement in mAP score of about 12%.
Predicting movement of objects while the action of learning agent interacts with the dynamics of the scene still remains a key challenge in robotics. We propose a multi-layer Long Short Term Memory (LSTM) autoendocer network that predicts future frames for a robot navigating in a dynamic environment with moving obstacles. The autoencoder network is composed of a state and action conditioned decoder network that reconstructs the future frames of video, conditioned on the action taken by the agent. The input image frames are first transformed into low dimensional feature vectors with a pre-trained encoder network and then reconstructed with the LSTM autoencoder network to generate the future frames. A virtual environment, based on the OpenAi-Gym framework for robotics, is used to gather training data and test the proposed network. The initial experiments show promising results indicating that these predicted frames can be used by an appropriate reinforcement learning framework in future to navigate around dynamic obstacles.
We present Piano Genie, an intelligent controller which allows non-musicians to improvise on the piano. With Piano Genie, a user performs on a simple interface with eight buttons, and their performance is decoded into the space of plausible piano music in real time. To learn a suitable mapping procedure for this problem, we train recurrent neural network autoencoders with discrete bottlenecks: an encoder learns an appropriate sequence of buttons corresponding to a piano piece, and a decoder learns to map this sequence back to the original piece. During performance, we substitute a user's input for the encoder output, and play the decoder's prediction each time the user presses a button. To improve the intuitiveness of Piano Genie's performance behavior, we impose musically meaningful constraints over the encoder's outputs.
Different texts shall by nature correspond to different number of keyphrases. This desideratum is largely missing from existing neural keyphrase generation models. In this study, we address this problem from both modeling and evaluation perspectives. We first propose a recurrent generative model that generates multiple keyphrases as delimiter-separated sequences. Generation diversity is further enhanced with two novel techniques by manipulating decoder hidden states. In contrast to previous approaches, our model is capable of generating diverse keyphrases and controlling number of outputs. We further propose two evaluation metrics tailored towards the variable-number generation. We also introduce a new dataset StackEx that expands beyond the only existing genre (i.e., academic writing) in keyphrase generation tasks. With both previous and new evaluation metrics, our model outperforms strong baselines on all datasets.
Most existing studies in text-to-SQL tasks do not require generating complex SQL queries with multiple clauses or sub-queries, and generalizing to new, unseen databases. In this paper we propose SyntaxSQLNet, a syntax tree network to address the complex and cross-domain text-to-SQL generation task. SyntaxSQLNet employs a SQL specific syntax tree-based decoder with SQL generation path history and table-aware column attention encoders. We evaluate SyntaxSQLNet on the Spider text-to-SQL task, which contains databases with multiple tables and complex SQL queries with multiple SQL clauses and nested queries. We use a database split setting where databases in the test set are unseen during training. Experimental results show that SyntaxSQLNet can handle a significantly greater number of complex SQL examples than prior work, outperforming the previous state-of-the-art model by 7.3% in exact matching accuracy. We also show that SyntaxSQLNet can further improve the performance by an additional 7.5% using a cross-domain augmentation method, resulting in a 14.8% improvement in total. To our knowledge, we are the first to study this complex and cross-domain text-to-SQL task.