Attention-based encoder-decoder neural network models have recently shown promising results in goal-oriented dialogue systems. However, these models struggle to reason over and incorporate state-full knowledge while preserving their end-to-end text generation functionality. Since such models can greatly benefit from user intent and knowledge graph integration, in this paper we propose an RNN-based end-to-end encoder-decoder architecture which is trained with joint embeddings of the knowledge graph and the corpus as input. The model provides an additional integration of user intent along with text generation, trained with a multi-task learning paradigm along with an additional regularization technique to penalize generating the wrong entity as output. The model further incorporates a Knowledge Graph entity lookup during inference to guarantee the generated output is state-full based on the local knowledge graph provided. We finally evaluated the model using the BLEU score, empirical evaluation depicts that our proposed architecture can aid in the betterment of task-oriented dialogue system`s performance.
We propose smoothed max pooling loss and its application to keyword spotting systems. The proposed approach jointly trains an encoder (to detect keyword parts) and a decoder (to detect whole keyword) in a semi-supervised manner. The proposed new loss function allows training a model to detect parts and whole of a keyword, without strictly depending on frame-level labeling from LVCSR (Large vocabulary continuous speech recognition), making further optimization possible. The proposed system outperforms the baseline keyword spotting model in [1] due to increased optimizability. Further, it can be more easily adapted for on-device learning applications due to reduced dependency on LVCSR.
We design an online end-to-end speech recognition system based on Time-Depth Separable (TDS) convolutions and Connectionist Temporal Classification (CTC). We improve the core TDS architecture in order to limit the future context and hence reduce latency while maintaining accuracy. The system has almost three times the throughput of a well tuned hybrid ASR baseline while also having lower latency and a better word error rate. Also important to the efficiency of the recognizer is our highly optimized beam search decoder. To show the impact of our design choices, we analyze throughput, latency, accuracy, and discuss how these metrics can be tuned based on the user requirements.
In the present paper, we propose the model of {\it structural information learning machines} (SiLeM for short), leading to a mathematical definition of learning by merging the theories of computation and information. Our model shows that the essence of learning is {\it to gain information}, that to gain information is {\it to eliminate uncertainty} embedded in a data space, and that to eliminate uncertainty of a data space can be reduced to an optimization problem, that is, an {\it information optimization problem}, which can be realized by a general {\it encoding tree method}. The principle and criterion of the structural information learning machines are maximization of {\it decoding information} from the data points observed together with the relationships among the data points, and semantical {\it interpretation} of syntactical {\it essential structure}, respectively. A SiLeM machine learns the laws or rules of nature. It observes the data points of real world, builds the {\it connections} among the observed data and constructs a {\it data space}, for which the principle is to choose the way of connections of data points so that the {\it decoding information} of the data space is maximized, finds the {\it encoding tree} of the data space that minimizes the dynamical uncertainty of the data space, in which the encoding tree is hence referred to as a {\it decoder}, due to the fact that it has already eliminated the maximum amount of uncertainty embedded in the data space, interprets the {\it semantics} of the decoder, an encoding tree, to form a {\it knowledge tree}, extracts the {\it remarkable common features} for both semantical and syntactical features of the modules decoded by a decoder to construct {\it trees of abstractions}, providing the foundations for {\it intuitive reasoning} in the learning when new data are observed.
Objective : Abdominal anatomy segmentation is crucial for numerous applications from computer-assisted diagnosis to image-guided surgery. In this context, we address fully-automated multi-organ segmentation from abdominal CT and MR images using deep learning. Methods: The proposed model extends standard conditional generative adversarial networks. Additionally to the discriminator which enforces the model to create realistic organ delineations, it embeds cascaded partially pre-trained convolutional encoder-decoders as generator. Encoder fine-tuning from a large amount of non-medical images alleviates data scarcity limitations. The network is trained end-to-end to benefit from simultaneous multi-level segmentation refinements using auto-context. Results : Employed for healthy liver, kidneys and spleen segmentation, our pipeline provides promising results by outperforming state-of-the-art encoder-decoder schemes. Followed for the Combined Healthy Abdominal Organ Segmentation (CHAOS) challenge organized in conjunction with the IEEE International Symposium on Biomedical Imaging 2019, it gave us the first rank for three competition categories: liver CT, liver MR and multi-organ MR segmentation. Conclusion : Combining cascaded convolutional and adversarial networks strengthens the ability of deep learning pipelines to automatically delineate multiple abdominal organs, with good generalization capability. Significance : The comprehensive evaluation provided suggests that better guidance could be achieved to help clinicians in abdominal image interpretation and clinical decision making.
Recently many works attempt to develop image compression models based on deep learning architectures, where the uniform scalar quantizer (SQ) is commonly applied to the feature maps between the encoder and decoder. In this paper, we propose to incorporate trellis coded quantizer (TCQ) into a deep learning based image compression framework. A soft-to-hard strategy is applied to allow for back propagation during training. We develop a simple image compression model that consists of three subnetworks (encoder, decoder and entropy estimation), and optimize all of the components in an end-to-end manner. We experiment on two high resolution image datasets and both show that our model can achieve superior performance at low bit rates. We also show the comparisons between TCQ and SQ based on our proposed baseline model and demonstrate the advantage of TCQ.
Scene Text Recognition is a challenging problem because of irregular styles and various distortions. This paper proposed an end-to-end trainable model consists of a finer rectification module and a bidirectional attentional recognition network(Firbarn). The rectification module adopts finer grid to rectify the distorted input image and the bidirectional decoder contains only one decoding layer instead of two separated one. Firbarn can be trained in a weak supervised way, only requiring the scene text images and the corresponding word labels. With the flexible rectification and the novel bidirectional decoder, the results of extensive evaluation on the standard benchmarks show Firbarn outperforms previous works, especially on irregular datasets.
Generative models, such as GANs, learn an explicit low-dimensional representation of a particular class of images, and so they may be used as natural image priors for solving inverse problems such as image restoration and compressive sensing. GAN priors have demonstrated impressive performance on these tasks, but they can exhibit substantial representation error for both in-distribution and out-of-distribution images, because of the mismatch between the learned, approximate image distribution and the data generating distribution. In this paper, we demonstrate a method for reducing the representation error of GAN priors by modeling images as the linear combination of a GAN prior with a Deep Decoder. The deep decoder is an underparameterized and most importantly unlearned natural signal model similar to the Deep Image Prior. No knowledge of the specific inverse problem is needed in the training of the GAN underlying our method. For compressive sensing and image superresolution, our hybrid model exhibits consistently higher PSNRs than both the GAN priors and Deep Decoder separately, both on in-distribution and out-of-distribution images. This model provides a method for extensibly and cheaply leveraging both the benefits of learned and unlearned image recovery priors in inverse problems.
This paper proposes a novel deep subspace clustering approach which uses convolutional autoencoders to transform input images into new representations lying on a union of linear subspaces. The first contribution of our work is to insert multiple fully-connected linear layers between the encoder layers and their corresponding decoder layers to promote learning more favorable representations for subspace clustering. These connection layers facilitate the feature learning procedure by combining low-level and high-level information for generating multiple sets of self-expressive and informative representations at different levels of the encoder. Moreover, we introduce a novel loss minimization problem which leverages an initial clustering of the samples to effectively fuse the multi-level representations and recover the underlying subspaces more accurately. The loss function is then minimized through an iterative scheme which alternatively updates the network parameters and produces new clusterings of the samples. Experiments on four real-world datasets demonstrate that our approach exhibits superior performance compared to the state-of-the-art methods on most of the subspace clustering problems.
This paper addresses two crucial problems of learning disentangled image representations, namely controlling the degree of disentanglement during image editing, and balancing the disentanglement strength and the reconstruction quality. To encourage disentanglement, we devise a distance covariance based decorrelation regularization. Further, for the reconstruction step, our model leverages a soft target representation combined with the latent image code. By exploring the real-valued space of the soft target representation, we are able to synthesize novel images with the designated properties. To improve the perceptual quality of images generated by autoencoder (AE)-based models, we extend the encoder-decoder architecture with the generative adversarial network (GAN) by collapsing the AE decoder and the GAN generator into one. We also design a classification based protocol to quantitatively evaluate the disentanglement strength of our model. Experimental results showcase the benefits of the proposed model.
For a robot to perform complex manipulation tasks, it is necessary for it to have a good grasping ability. However, vision based robotic grasp detection is hindered by the unavailability of sufficient labelled data. Furthermore, the application of semi-supervised learning techniques to grasp detection is under-explored. In this paper, a semi-supervised learning based grasp detection approach has been presented, which models a discrete latent space using a Vector Quantized Variational AutoEncoder (VQ-VAE). To the best of our knowledge, this is the first time a Variational AutoEncoder (VAE) has been applied in the domain of robotic grasp detection. The VAE helps the model in generalizing beyond the Cornell Grasping Dataset (CGD) despite having a limited amount of labelled data by also utilizing the unlabelled data. This claim has been validated by testing the model on images, which are not available in the CGD. Along with this, we augment the Generative Grasping Convolutional Neural Network (GGCNN) architecture with the decoder structure used in the VQ-VAE model with the intuition that it should help to regress in the vector-quantized latent space. Subsequently, the model performs significantly better than the existing approaches which do not make use of unlabelled images to improve the grasp.
Recently, Transformer has gained success in automatic speech recognition (ASR) field. However, it is challenging to deploy a Transformer-based end-to-end (E2E) model for online speech recognition. In this paper, we propose the Transformer-based online CTC/attention E2E ASR architecture, which contains the chunk self-attention encoder (chunk-SAE) and the monotonic truncated attention (MTA) based self-attention decoder (SAD). Firstly, the chunk-SAE splits the speech into isolated chunks. To reduce the computational cost and improve the performance, we propose the state reuse chunk-SAE. Sencondly, the MTA based SAD truncates the speech features monotonically and performs attention on the truncated features. To support the online recognition, we integrate the state reuse chunk-SAE and the MTA based SAD into online CTC/attention architecture. We evaluate the proposed online models on the HKUST Mandarin ASR benchmark and achieve a 23.66% character error rate (CER) with a 320 ms latency. Our online model yields as little as 0.19% absolute CER degradation compared with the offline baseline, and achieves significant improvement over our prior work on Long Short-Term Memory (LSTM) based online E2E models.
This paper demonstrates that multilingual denoising pre-training produces significant performance gains across a wide variety of machine translation (MT) tasks. We present mBART -- a sequence-to-sequence denoising auto-encoder pre-trained on large-scale monolingual corpora in many languages using the BART objective. mBART is one of the first methods for pre-training a complete sequence-to-sequence model by denoising full texts in multiple languages, while previous approaches have focused only on the encoder, decoder, or reconstructing parts of the text. Pre-training a complete model allows it to be directly fine tuned for supervised (both sentence-level and document-level) and unsupervised machine translation, with no task-specific modifications. We demonstrate that adding mBART initialization produces performance gains in all but the highest-resource settings, including up to 12 BLEU points for low resource MT and over 5 BLEU points for many document-level and unsupervised models. We also show it also enables new types of transfer to language pairs with no bi-text or that were not in the pre-training corpus, and present extensive analysis of which factors contribute the most to effective pre-training.
This paper proposes a human-aware deblurring model that disentangles the motion blur between foreground (FG) humans and background (BG). The proposed model is based on a triple-branch encoder-decoder architecture. The first two branches are learned for sharpening FG humans and BG details, respectively; while the third one produces global, harmonious results by comprehensively fusing multi-scale deblurring information from the two domains. The proposed model is further endowed with a supervised, human-aware attention mechanism in an end-to-end fashion. It learns a soft mask that encodes FG human information and explicitly drives the FG/BG decoder-branches to focus on their specific domains. To further benefit the research towards Human-aware Image Deblurring, we introduce a large-scale dataset, named HIDE, which consists of 8,422 blurry and sharp image pairs with 65,784 densely annotated FG human bounding boxes. HIDE is specifically built to span a broad range of scenes, human object sizes, motion patterns, and background complexities. Extensive experiments on public benchmarks and our dataset demonstrate that our model performs favorably against the state-of-the-art motion deblurring methods, especially in capturing semantic details.
The accurate retinal vessel segmentation (RVS) is of great significance to assist doctors in the diagnosis of ophthalmology diseases and other systemic diseases. Manually designing a valid neural network architecture for retinal vessel segmentation requires high expertise and a large workload. In order to improve the performance of vessel segmentation and reduce the workload of manually designing neural network, we propose novel approach which applies neural architecture search (NAS) to optimize an encoder-decoder architecture for retinal vessel segmentation. A modified evolutionary algorithm is used to evolve the architectures of encoder-decoder framework with limited computing resources. The evolved model obtained by the proposed approach achieves top performance among all compared methods on the three datasets, namely DRIVE, STARE and CHASE_DB1, but with much fewer parameters. Moreover, the results of cross-training show that the evolved model is with considerable scalability, which indicates a great potential for clinical disease diagnosis.
Conditional image modeling based on textual descriptions is a relatively new domain in unsupervised learning. Previous approaches use a latent variable model and generative adversarial networks. While the formers are approximated by using variational auto-encoders and rely on the intractable inference that can hamper their performance, the latter is unstable to train due to Nash equilibrium based objective function. We develop a tractable and stable caption-based image generation model. The model uses an attention-based encoder to learn word-to-pixel dependencies. A conditional autoregressive based decoder is used for learning pixel-to-pixel dependencies and generating images. Experimentations are performed on Microsoft COCO, and MNIST-with-captions datasets and performance is evaluated by using the Structural Similarity Index. Results show that the proposed model performs better than contemporary approaches and generate better quality images. Keywords: Generative image modeling, autoregressive image modeling, caption-based image generation, neural attention, recurrent neural networks.
Neural conversation systems generate responses based on the sequence-to-sequence (SEQ2SEQ) paradigm. Typically, the model is equipped with a single set of learned parameters to generate responses for given input contexts. When confronting diverse conversations, its adaptability is rather limited and the model is hence prone to generate generic responses. In this work, we propose an {\bf Ada}ptive {\bf N}eural {\bf D}ialogue generation model, \textsc{AdaND}, which manages various conversations with conversation-specific parameterization. For each conversation, the model generates parameters of the encoder-decoder by referring to the input context. In particular, we propose two adaptive parameterization mechanisms: a context-aware and a topic-aware parameterization mechanism. The context-aware parameterization directly generates the parameters by capturing local semantics of the given context. The topic-aware parameterization enables parameter sharing among conversations with similar topics by first inferring the latent topics of the given context and then generating the parameters with respect to the distributional topics. Extensive experiments conducted on a large-scale real-world conversational dataset show that our model achieves superior performance in terms of both quantitative metrics and human evaluations.
We consider near maximum-likelihood (ML) decoding of short linear block codes based on neural belief propagation (BP) decoding recently introduced by Nachmani et al.. While this method significantly outperforms conventional BP decoding, the underlying parity-check matrix may still limit the overall performance. In this paper, we introduce a method to tailor an overcomplete parity-check matrix to (neural) BP decoding using machine learning. We consider the weights in the Tanner graph as an indication of the importance of the connected check nodes (CNs) to decoding and use them to prune unimportant CNs. As the pruning is not tied over iterations, the final decoder uses a different parity-check matrix in each iteration. For Reed-Muller and short low-density parity-check codes, we achieve performance within 0.27 dB and 1.5 dB of the ML performance while reducing the complexity of the decoder.
We propose a new length-controllable abstractive summarization model. Recent state-of-the-art abstractive summarization models based on encoder-decoder models generate only one summary per source text. However, controllable summarization, especially of the length, is an important aspect for practical applications. Previous studies on length-controllable abstractive summarization incorporate length embeddings in the decoder module for controlling the summary length. Although the length embeddings can control where to stop decoding, they do not decide which information should be included in the summary within the length constraint. Unlike the previous models, our length-controllable abstractive summarization model incorporates a word-level extractive module in the encoder-decoder model instead of length embeddings. Our model generates a summary in two steps. First, our word-level extractor extracts a sequence of important words (we call it the "prototype text") from the source text according to the word-level importance scores and the length constraint. Second, the prototype text is used as additional input to the encoder-decoder model, which generates a summary by jointly encoding and copying words from both the prototype text and source text. Since the prototype text is a guide to both the content and length of the summary, our model can generate an informative and length-controlled summary. Experiments with the CNN/Daily Mail dataset and the NEWSROOM dataset show that our model outperformed previous models in length-controlled settings.
Analyzing the urban trajectory in cities has become an important topic in data mining. How can we model the human mobility consisting of stay and travel from the raw trajectory data? How can we infer such a mobility model from the single trajectory information? How can we further generalize the mobility inference to accommodate the real-world trajectory data that is sparsely sampled over time? In this paper, based on formal and rigid definitions of the stay/travel mobility, we propose a single trajectory inference algorithm that utilizes a generic long-tailed sparsity pattern in the large-scale trajectory data. The algorithm guarantees a 100\% precision in the stay/travel inference with a provable lower-bound in the recall. Furthermore, we introduce an encoder-decoder learning architecture that admits multiple trajectories as inputs. The architecture is optimized for the mobility inference problem through customized embedding and learning mechanism. Evaluations with three trajectory data sets of 40 million urban users validate the performance guarantees of the proposed inference algorithm and demonstrate the superiority of our deep learning model, in comparison to well-known sequence learning methods. On extremely sparse trajectories, the deep learning model achieves a 2$\times$ overall accuracy improvement from the single trajectory inference algorithm, through proven scalability and generalizability to large-scale versatile training data.