Embedding acoustic information into fixed length representations is of interest for a whole range of applications in speech and audio technology. Two novel unsupervised approaches to generate acoustic embeddings by modelling of acoustic context are proposed. The first approach is a contextual joint factor synthesis encoder, where the encoder in an encoder/decoder framework is trained to extract joint factors from surrounding audio frames to best generate the target output. The second approach is a contextual joint factor analysis encoder, where the encoder is trained to analyse joint factors from the source signal that correlates best with the neighbouring audio. To evaluate the effectiveness of our approaches compared to prior work, two tasks are conducted -- phone classification and speaker recognition -- and test on different TIMIT data sets. Experimental results show that one of the proposed approaches outperforms phone classification baselines, yielding a classification accuracy of 74.1%. When using additional out-of-domain data for training, an additional 3% improvements can be obtained, for both for phone classification and speaker recognition tasks.
In this work, we study how the finetuning stage in the pretrain-finetune framework changes the behavior of a pretrained neural language generator. We focus on the transformer encoder-decoder model for the open-domain dialogue response generation task. Our major finding is that after standard finetuning, the model forgets some of the important language generation skills acquired during large-scale pretraining. We demonstrate the forgetting phenomenon through a set of detailed behavior analysis from the perspectives of knowledge transfer, context sensitivity, and function space projection. As a preliminary attempt to alleviate the forgetting problem, we propose an intuitive finetuning strategy named "mix-review". We find that mix-review effectively regularizes the finetuning process, and the forgetting problem is alleviated to some extent. Finally, we discuss interesting behavior of the resulting dialogue model and its implications.
This paper describes the Imperial College London team's submission to the 2019' VATEX video captioning challenge, where we first explore two sequence-to-sequence models, namely a recurrent (GRU) model and a transformer model, which generate captions from the I3D action features. We then investigate the effect of dropping the encoder and the attention mechanism and instead conditioning the GRU decoder over two different vectorial representations: (i) a max-pooled action feature vector and (ii) the output of a multi-label classifier trained to predict visual entities from the action features. Our baselines achieved scores comparable to the official baseline. Conditioning over entity predictions performed substantially better than conditioning on the max-pooled feature vector, and only marginally worse than the GRU-based sequence-to-sequence baseline.
Domain adaptation in natural language generation (NLG) remains challenging because of the high complexity of input semantics across domains and limited data of a target domain. This is particularly the case for dialogue systems, where we want to be able to seamlessly include new domains into the conversation. Therefore, it is crucial for generation models to share knowledge across domains for the effective adaptation from one domain to another. In this study, we exploit a tree-structured semantic encoder to capture the internal structure of complex semantic representations required for multi-domain dialogues in order to facilitate knowledge sharing across domains. In addition, a layer-wise attention mechanism between the tree encoder and the decoder is adopted to further improve the model's capability. The automatic evaluation results show that our model outperforms previous methods in terms of the BLEU score and the slot error rate, in particular when the adaptation data is limited. In subjective evaluation, human judges tend to prefer the sentences generated by our model, rating them more highly on informativeness and naturalness than other systems.
In this paper, we investigate a novel problem of telling the difference between image pairs in natural language. Compared to previous approaches for single image captioning, it is challenging to fetch linguistic representation from two independent visual information. To this end, we have proposed an effective encoder-decoder caption framework based on Hyper Convolution Net. In addition, a series of novel feature fusing techniques for pairwise visual information fusing are introduced and a discriminating referee is proposed to evaluate the pipeline. Because of the lack of appropriate datasets to support this task, we have collected and annotated a large new dataset with Amazon Mechanical Turk (AMT) for generating captions in a pairwise manner (with 14764 images and 26710 image pairs in total). The dataset is the first one on the relative difference caption task that provides descriptions in free language. We evaluate the effectiveness of our model on two datasets in the field and it outperforms the state-of-the-art approach by a large margin.
Language Identification (LID) is a challenging task, especially when the input texts are short and noisy such as posts and statuses on social media or chat logs on gaming forums. The task has been tackled by either designing a feature set for a traditional classifier (e.g. Naive Bayes) or applying a deep neural network classifier (e.g. Bi-directional Gated Recurrent Unit, Encoder-Decoder). These methods are usually trained and tested on a huge amount of private data, then used and evaluated as off-the-shelf packages by other researchers using their own datasets, and consequently the various results published are not directly comparable. In this paper, we first create a new massive labelled dataset based on one year of Twitter data. We use this dataset to test several existing language identification systems, in order to obtain a set of coherent benchmarks, and we make our dataset publicly available so that others can add to this set of benchmarks. Finally, we propose a shallow but efficient neural LID system, which is a ngram-regional convolution neural network enhanced with an attention mechanism. Experimental results show that our architecture is able to predict tens of thousands of samples per second and surpasses all state-of-the-art systems with an improvement of 5%.
Depth completion aims to recover dense depth maps from sparse depth measurements. It is of increasing importance for autonomous driving and draws increasing attention from the vision community. Most of existing methods directly train a network to learn a mapping from sparse depth inputs to dense depth maps, which has difficulties in utilizing the 3D geometric constraints and handling the practical sensor noises. In this paper, to regularize the depth completion and improve the robustness against noise, we propose a unified CNN framework that 1) models the geometric constraints between depth and surface normal in a diffusion module and 2) predicts the confidence of sparse LiDAR measurements to mitigate the impact of noise. Specifically, our encoder-decoder backbone predicts surface normals, coarse depth and confidence of LiDAR inputs simultaneously, which are subsequently inputted into our diffusion refinement module to obtain the final completion results. Extensive experiments on KITTI depth completion dataset and NYU-Depth-V2 dataset demonstrate that our method achieves state-of-the-art performance. Further ablation study and analysis give more insights into the proposed method and demonstrate the generalization capability and stability of our model.
Image captioning is a research hotspot where encoder-decoder models combining convolutional neural network (CNN) and long short-term memory (LSTM) achieve promising results. Despite significant progress, these models generate sentences differently from human cognitive styles. Existing models often generate a complete sentence from the first word to the end, without considering the influence of the following words on the whole sentence generation. In this paper, we explore the utilization of a human-like cognitive style, i.e., building overall cognition for the image to be described and the sentence to be constructed, for enhancing computer image understanding. This paper first proposes a Mutual-aid network structure with Bidirectional LSTMs (MaBi-LSTMs) for acquiring overall contextual information. In the training process, the forward and backward LSTMs encode the succeeding and preceding words into their respective hidden states by simultaneously constructing the whole sentence in a complementary manner. In the captioning process, the LSTM implicitly utilizes the subsequent semantic information contained in its hidden states. In fact, MaBi-LSTMs can generate two sentences in forward and backward directions. To bridge the gap between cross-domain models and generate a sentence with higher quality, we further develop a cross-modal attention mechanism to retouch the two sentences by fusing their salient parts as well as the salient areas of the image. Experimental results on the Microsoft COCO dataset show that the proposed model improves the performance of encoder-decoder models and achieves state-of-the-art results.
In order to efficiently transmit and store speech signals, speech codecs create a minimally redundant representation of the input signal which is then decoded at the receiver with the best possible perceptual quality. In this work we demonstrate that a neural network architecture based on VQ-VAE with a WaveNet decoder can be used to perform very low bit-rate speech coding with high reconstruction quality. A prosody-transparent and speaker-independent model trained on the LibriSpeech corpus coding audio at 1.6 kbps exhibits perceptual quality which is around halfway between the MELP codec at 2.4 kbps and AMR-WB codec at 23.05 kbps. In addition, when training on high-quality recorded speech with the test speaker included in the training set, a model coding speech at 1.6 kbps produces output of similar perceptual quality to that generated by AMR-WB at 23.05 kbps.
Distributed representations have been used to support downstream tasks in healthcare recently. Healthcare data (e.g., electronic health records) contain multiple modalities of data from heterogeneous sources that can provide complementary information, alongside an added dimension to learning personalized patient representations. To this end, in this paper we propose a novel unsupervised encoder-decoder model, namely Mixed Pooling Multi-View Attention Autoencoder (MPVAA), that generates patient representations encapsulating a holistic view of their medical profile. Specifically, by first learning personalized graph embeddings pertaining to each patient's heterogeneous healthcare data, it then integrates the non-linear relationships among them into a unified representation through multi-view attention mechanism. Additionally, a mixed pooling strategy is incorporated in the encoding step to learn diverse information specific to each data modality. Experiments conducted for multiple tasks demonstrate the effectiveness of the proposed model over the state-of-the-art representation learning methods in healthcare.
Automatic text summarization (ATS) has recently achieved impressive performance thanks to recent advances in deep learning and the availability of large-scale corpora. To make the summarization results more faithful, this paper presents an unsupervised approach that combines rhetorical structure theory, deep neural model and domain knowledge concern for ATS. This architecture mainly contains three components: domain knowledge base construction based on representation learning, attentional encoder-decoder model for rhetorical parsing and subroutine-based model for text summarization. Domain knowledge can be effectively used for unsupervised rhetorical parsing thus rhetorical structure trees for each document can be derived. In the unsupervised rhetorical parsing module, the idea of translation was adopted to alleviate the problem of data scarcity. The subroutine-based summarization model purely depends on the derived rhetorical structure trees and can generate content-balanced results. To evaluate the summary results without golden standard, we proposed an unsupervised evaluation metric, whose hyper-parameters were tuned by supervised learning. Experimental results show that, on a large-scale Chinese dataset, our proposed approach can obtain comparable performances compared with existing methods.
We evaluate three simple, normalization-centric changes to improve Transformer training. First, we show that pre-norm residual connections (PreNorm) and smaller initializations enable warmup-free, validation-based training with large learning rates. Second, we propose $\ell_2$ normalization with a single scale parameter (ScaleNorm) for faster training and better performance. Finally, we reaffirm the effectiveness of normalizing word embeddings to a fixed length (FixNorm). On five low-resource translation pairs from TED Talks-based corpora, these changes always converge, giving an average +1.1 BLEU over state-of-the-art bilingual baselines and a new 32.8 BLEU on IWSLT'15 English-Vietnamese. We observe sharper performance curves, more consistent gradient norms, and a linear relationship between activation scaling and decoder depth. Surprisingly, in the high-resource setting (WMT'14 English-German), ScaleNorm and FixNorm remain competitive but PreNorm degrades performance.
Subspace clustering aims to cluster unlabeled data that lies in a union of low-dimensional linear subspaces. Deep subspace clustering approaches based on auto-encoders have become very popular to solve subspace clustering problems. However, the training of current deep methods converges slowly, which is much less efficient than traditional approaches. We propose a Residual Encoder-Decoder network for deep Subspace Clustering (RED-SC), which symmetrically links convolutional and deconvolutional layers with skip-layer connections, with which the training converges much faster. We use a self-expressive layer to generate more accurate linear representation coefficients through different latent representations from multiple latent spaces. Experiments show the superiority of RED-SC in training efficiency and clustering accuracy. Moreover, we are the first one to apply residual encoder-decoder on unsupervised learning tasks.
With the advancement of remote-sensed imaging large volumes of very high resolution land cover images can now be obtained. Automation of object recognition in these 2D images, however, is still a key issue. High intra-class variance and low inter-class variance in Very High Resolution (VHR) images hamper the accuracy of prediction in object recognition tasks. Most successful techniques in various computer vision tasks recently are based on deep supervised learning. In this work, a deep Convolutional Neural Network (CNN) based on symmetric encoder-decoder architecture with skip connections is employed for the 2D semantic segmentation of most common land cover object classes - impervious surface, buildings, low vegetation, trees and cars. Atrous convolutions are employed to have large receptive field in the proposed CNN model. Further, the CNN outputs are post-processed using Fully Connected Conditional Random Field (FCRF) model to refine the CNN pixel label predictions. The proposed CNN-FCRF model achieves an overall accuracy of 90.5% on the ISPRS Vaihingen Dataset.
Extractive keyphrase generation research has been around since the nineties, but the more advanced abstractive approach based on the encoder-decoder framework and sequence-to-sequence learning has been explored only recently. In fact, more than a dozen of abstractive methods have been proposed in the last three years, producing meaningful keyphrases and achieving state-of-the-art scores. In this survey, we examine various aspects of the extractive keyphrase generation methods and focus mostly on the more recent abstractive methods that are based on neural networks. We pay particular attention to the mechanisms that have driven the perfection of the later. A huge collection of scientific article metadata and the corresponding keyphrases is created and released for the research community. We also present various keyphrase generation and text summarization research patterns and trends of the last two decades.
Bit-depth is the number of bits for each color channel of a pixel in an image. Although many modern displays support unprecedented higher bit-depth to show more realistic and natural colors with a high dynamic range, most media sources are still in bit-depth of 8 or lower. Since insufficient bit-depth may generate annoying false contours or lose detailed visual appearance, bit-depth expansion (BDE) from low bit-depth (LBD) images to high bit-depth (HBD) images becomes more and more important. In this paper, we adopt a learning-based approach for BDE and propose a novel CNN-based bit-depth expansion network (BitNet) that can effectively remove false contours and restore visual details at the same time. We have carefully designed our BitNet based on an encoder-decoder architecture with dilated convolutions and a novel multi-scale feature integration. We have performed various experiments with four different datasets including MIT-Adobe FiveK, Kodak, ESPL v2, and TESTIMAGES, and our proposed BitNet has achieved state-of-the-art performance in terms of PSNR and SSIM among other existing BDE methods and famous CNN-based image processing networks. Unlike previous methods that separately process each color channel, we treat all RGB channels at once and have greatly improved color restoration. In addition, our network has shown the fastest computational speed in near real-time.
We present Panoptic-DeepLab, a bottom-up and single-shot approach for panoptic segmentation. Our Panoptic-DeepLab is conceptually simple and delivers state-of-the-art results. In particular, we adopt the dual-ASPP and dual-decoder structures specific to semantic, and instance segmentation, respectively. The semantic segmentation branch is the same as the typical design of any semantic segmentation model (e.g., DeepLab), while the instance segmentation branch is class-agnostic, involving a simple instance center regression. Our single Panoptic-DeepLab sets the new state-of-art at all three Cityscapes benchmarks, reaching 84.2% mIoU, 39.0% AP, and 65.5% PQ on test set, and advances results on the other challenging Mapillary Vistas.
We introduce the Mutual Information Machine (MIM), a novel formulation of representation learning, using a joint distribution over the observations and latent state in an encoder/decoder framework. Our key principles are symmetry and mutual information, where symmetry encourages the encoder and decoder to learn different factorizations of the same underlying distribution, and mutual information, to encourage the learning of useful representations for downstream tasks. Our starting point is the symmetric Jensen-Shannon divergence between the encoding and decoding joint distributions, plus a mutual information encouraging regularizer. We show that this can be bounded by a tractable cross entropy loss function between the true model and a parameterized approximation, and relate this to the maximum likelihood framework. We also relate MIM to variational autoencoders (VAEs) and demonstrate that MIM is capable of learning symmetric factorizations, with high mutual information that avoids posterior collapse.
Data availability is a bottleneck during early stages of development of new capabilities for intelligent artificial agents. We investigate the use of text generation techniques to augment the training data of a popular commercial artificial agent across categories of functionality, with the goal of faster development of new functionality. We explore a variety of encoder-decoder generative models for synthetic training data generation and propose using conditional variational auto-encoders. Our approach requires only direct optimization, works well with limited data and significantly outperforms the previous controlled text generation techniques. Further, the generated data are used as additional training samples in an extrinsic intent classification task, leading to improved performance by up to 5\% absolute f-score in low-resource cases, validating the usefulness of our approach.
Generative seq2seq dialogue systems are trained to predict the next word in dialogues that have already occurred. They can learn from large unlabeled conversation datasets, build a deep understanding of conversational context, and generate a wide variety of responses. This flexibility comes at the cost of control. Undesirable responses in the training data will be reproduced by the model at inference time, and longer generations often don't make sense. Instead of generating responses one word at a time, we train a classifier to choose from a predefined list of full responses. The classifier is trained on (conversation context, response class) pairs, where each response class is a noisily labeled group of interchangeable responses. At inference, we generate the exemplar response associated with the predicted response class. Experts can edit and improve these exemplar responses over time without retraining the classifier or invalidating old training data. Human evaluation of 775 unseen doctor/patient conversations shows that this tradeoff improves responses. Only 12% of our discriminative approach's responses are worse than the doctor's response in the same conversational context, compared to 18% for the generative model. A discriminative model trained without any manual labeling of response classes achieves equal performance to the generative model.