Non-autoregressive mechanisms can significantly decrease inference time for speech transformers, especially when the single step variant is applied. Previous work on CTC alignment-based single step non-autoregressive transformer (CASS-NAT) has shown a large real time factor (RTF) improvement over autoregressive transformers (AT). In this work, we propose several methods to improve the accuracy of the end-to-end CASS-NAT, followed by performance analyses. First, convolution augmented self-attention blocks are applied to both the encoder and decoder modules. Second, we propose to expand the trigger mask (acoustic boundary) for each token to increase the robustness of CTC alignments. In addition, iterated loss functions are used to enhance the gradient update of low-layer parameters. Without using an external language model, the WERs of the improved CASS-NAT, when using the three methods, are 3.1%/7.2% on Librispeech test clean/other sets and the CER is 5.4% on the Aishell1 test set, achieving a 7%~21% relative WER/CER improvement. For the analyses, we plot attention weight distributions in the decoders to visualize the relationships between token-level acoustic embeddings. When the acoustic embeddings are visualized, we find that they have a similar behavior to word embeddings, which explains why the improved CASS-NAT performs similarly to AT.
High-resolution satellite imagery is critical for various earth observation applications related to environment monitoring, geoscience, forecasting, and land use analysis. However, the acquisition cost of such high-quality imagery due to the scarcity of providers and needs for high-frequency revisits restricts its accessibility in many fields. In this work, we present a data-driven, multi-image super resolution approach to alleviate these problems. Our approach is based on an end-to-end deep neural network that consists of an encoder, a fusion module, and a decoder. The encoder extracts co-registered highly efficient feature representations from low-resolution images of a scene. A Gated Re-current Unit (GRU)-based module acts as the fusion module, aggregating features into a combined representation. Finally, a decoder reconstructs the super-resolved image. The proposed model is evaluated on the PROBA-V dataset released in a recent competition held by the European Space Agency. Our results show that it performs among the top contenders and offers a new practical solution for real-world applications.
In this work we address the problem of landmark-based segmentation for anatomical structures. We propose HybridGNet, an encoder-decoder neural architecture which combines standard convolutions for image feature encoding, with graph convolutional neural networks to decode plausible representations of anatomical structures. We benchmark the proposed architecture considering other standard landmark and pixel-based models for anatomical segmentation in chest x-ray images, and found that HybridGNet is more robust to image occlusions. We also show that it can be used to construct landmark-based segmentations from pixel level annotations. Our experimental results suggest that HybridGNet produces accurate and anatomically plausible landmark-based segmentations, by naturally incorporating shape constraints within the decoding process via spectral convolutions.
The importance of building neural networks that can learn to reason has been well recognized in the neuro-symbolic community. In this paper, we apply neural pointer networks for conducting reasoning over symbolic knowledge bases. In doing so, we explore the benefits and limitations of encoder-decoder architectures in general and pointer networks in particular for developing accurate, generalizable and robust neuro-symbolic reasoners. Based on our experimental results, pointer networks performs remarkably well across multiple reasoning tasks while outperforming the previously reported state of the art by a significant margin. We observe that the Pointer Networks preserve their performance even when challenged with knowledge graphs of the domain/vocabulary it has never encountered before. To the best of our knowledge, this is the first study on neuro-symbolic reasoning using Pointer Networks. We hope our impressive results on these reasoning problems will encourage broader exploration of pointer networks' capabilities for reasoning over more complex logics and for other neuro-symbolic problems.
Deep complex convolution recurrent network (DCCRN), which extends CRN with complex structure, has achieved superior performance in MOS evaluation in Interspeech 2020 deep noise suppression challenge (DNS2020). This paper further extends DCCRN with the following significant revisions. We first extend the model to sub-band processing where the bands are split and merged by learnable neural network filters instead of engineered FIR filters, leading to a faster noise suppressor trained in an end-to-end manner. Then the LSTM is further substituted with a complex TF-LSTM to better model temporal dependencies along both time and frequency axes. Moreover, instead of simply concatenating the output of each encoder layer to the input of the corresponding decoder layer, we use convolution blocks to first aggregate essential information from the encoder output before feeding it to the decoder layers. We specifically formulate the decoder with an extra a priori SNR estimation module to maintain good speech quality while removing noise. Finally a post-processing module is adopted to further suppress the unnatural residual noise. The new model, named DCCRN+, has surpassed the original DCCRN as well as several competitive models in terms of PESQ and DNSMOS, and has achieved superior performance in the new Interspeech 2021 DNS challenge
Voice Conversion (VC) is a technique that aims to transform the non-linguistic information of a source utterance to change the perceived identity of the speaker. While there is a rich literature on VC, most proposed methods are trained and evaluated on clean speech recordings. However, many acoustic environments are noisy and reverberant, severely restricting the applicability of popular VC methods to such scenarios. To address this limitation, we propose Voicy, a new VC framework particularly tailored for noisy speech. Our method, which is inspired by the de-noising auto-encoders framework, is comprised of four encoders (speaker, content, phonetic and acoustic-ASR) and one decoder. Importantly, Voicy is capable of performing non-parallel zero-shot VC, an important requirement for any VC system that needs to work on speakers not seen during training. We have validated our approach using a noisy reverberant version of the LibriSpeech dataset. Experimental results show that Voicy outperforms other tested VC techniques in terms of naturalness and target speaker similarity in noisy reverberant environments.
Research in Curriculum Learning has shown better performance on the task by optimizing the sequence of the training data. Recent works have focused on using complex reinforcement learning techniques to find the optimal data ordering strategy to maximize learning for a given network. In this paper, we present a simple yet efficient technique based on continuous optimization trained with auto-encoding procedure. We call this new approach Training Sequence Optimization (TSO). With a usual encoder-decoder setup we try to learn the latent space continuous representation of the training strategy and a predictor network is used on the continuous representation to predict the accuracy of the strategy on the fixed network architecture. The performance predictor and encoder enable us to perform gradient-based optimization by gradually moving towards the latent space representation of training data ordering with potentially better accuracy. We show an empirical gain of 2AP with our generated optimal curriculum strategy over the random strategy using the CIFAR-100 and CIFAR-10 datasets and have better boosts than the existing state-of-the-art CL algorithms.
In design, fabrication, and control problems, we are often faced with the task of synthesis, in which we must generate an object or configuration that satisfies a set of constraints while maximizing one or more objective functions. The synthesis problem is typically characterized by a physical process in which many different realizations may achieve the goal. This many-to-one map presents challenges to the supervised learning of feed-forward synthesis, as the set of viable designs may have a complex structure. In addition, the non-differentiable nature of many physical simulations prevents efficient direct optimization. We address both of these problems with a two-stage neural network architecture that we may consider to be an autoencoder. We first learn the decoder: a differentiable surrogate that approximates the many-to-one physical realization process. We then learn the encoder, which maps from goal to design, while using the fixed decoder to evaluate the quality of the realization. We evaluate the approach on two case studies: extruder path planning in additive manufacturing and constrained soft robot inverse kinematics. We compare our approach to direct optimization of the design using the learned surrogate, and to supervised learning of the synthesis problem. We find that our approach produces higher quality solutions than supervised learning, while being competitive in quality with direct optimization, at a greatly reduced computational cost.
Automated Program Repair (APR) helps improve the efficiency of software development and maintenance. Recent APR techniques use deep learning, particularly the encoder-decoder architecture, to generate patches. Though existing DL-based APR approaches have proposed different encoder architectures, the decoder remains to be the standard one, which generates a sequence of tokens one by one to replace the faulty statement. This decoder has multiple limitations: 1) allowing to generate syntactically incorrect programs, 2) inefficiently representing small edits, and 3) not being able to generate project-specific identifiers. In this paper, we propose Recoder, a syntax-guided edit decoder with placeholder generation. Recoder is novel in multiple aspects: 1) Recoder generates edits rather than modified code, allowing efficient representation of small edits; 2) Recoder is syntax-guided, with the novel provider/decider architecture to ensure the syntactic correctness of the patched program and accurate generation; 3) Recoder generates placeholders that could be instantiated as project-specific identifiers later. We conduct experiments to evaluate Recoder on 395 bugs from Defects4J v1.2 and 420 additional bugs from Defects4J v2.0. Our results show that Recoder repairs 53 bugs on Defects4J v1.2, which achieves 21.4% improvement over the previous state-of-the-art approach for single-hunk bugs (TBar). Importantly, to our knowledge, Recoder is the first DL-based APR approach that has outperformed the traditional APR approaches on this dataset. Furthermore, Recoder also repairs 19 bugs on the additional bugs from Defects4J v2.0, which is 137.5% more than TBar (8 bugs) and 850% more than SimFix (2 bugs). This result suggests that Recoder has better generalizability than existing APR approaches.
The cyclically equivariant neural decoder was recently proposed in [Chen-Ye, International Conference on Machine Learning, 2021] to decode cyclic codes. In the same paper, a list decoding procedure was also introduced for two widely used classes of cyclic codes -- BCH codes and punctured Reed-Muller (RM) codes. While the list decoding procedure significantly improves the Frame Error Rate (FER) of the cyclically equivariant neural decoder, the Bit Error Rate (BER) of the list decoding procedure is even worse than the unique decoding algorithm when the list size is small. In this paper, we propose an improved version of the list decoding algorithm for BCH codes and punctured RM codes. Our new proposal significantly reduces the BER while maintaining the same (in some cases even smaller) FER. More specifically, our new decoder provides up to $2$dB gain over the previous list decoder when measured by BER, and the running time of our new decoder is $15\%$ smaller. Code available at https://github.com/improvedlistdecoder/code
We propose a simple yet effective method to compress an RNN-Transducer (RNN-T) through the well-known knowledge distillation paradigm. We show that the transducer's encoder outputs naturally have a high entropy and contain rich information about acoustically similar word-piece confusions. This rich information is suppressed when combined with the lower entropy decoder outputs to produce the joint network logits. Consequently, we introduce an auxiliary loss to distill the encoder logits from a teacher transducer's encoder, and explore training strategies where this encoder distillation works effectively. We find that tandem training of teacher and student encoders with an inplace encoder distillation outperforms the use of a pre-trained and static teacher transducer. We also report an interesting phenomenon we refer to as implicit distillation, that occurs when the teacher and student encoders share the same decoder. Our experiments show 5.37-8.4% relative word error rate reductions (WERR) on in-house test sets, and 5.05-6.18% relative WERRs on LibriSpeech test sets.
To overcome the limitations of convolutional neural network in the process of facial expression recognition, a facial expression recognition model Capsule-LSTM based on video frame sequence is proposed. This model is composed of three networks includingcapsule encoders, capsule decoders and LSTM network. The capsule encoder extracts the spatial information of facial expressions in video frames. Capsule decoder reconstructs the images to optimize the network. LSTM extracts the temporal information between video frames and analyzes the differences in expression changes between frames. The experimental results from the MMI dataset show that the Capsule-LSTM model proposed in this paper can effectively improve the accuracy of video expression recognition.
While multitask and transfer learning has shown to improve the performance of neural networks in limited data settings, they require pretraining of the model on large datasets beforehand. In this paper, we focus on improving the performance of weakly supervised sound event detection in low data and noisy settings simultaneously without requiring any pretraining task. To that extent, we propose a shared encoder architecture with sound event detection as a primary task and an additional secondary decoder for a self-supervised auxiliary task. We empirically evaluate the proposed framework for weakly supervised sound event detection on a remix dataset of the DCASE 2019 task 1 acoustic scene data with DCASE 2018 Task 2 sounds event data under 0, 10 and 20 dB SNR. To ensure we retain the localisation information of multiple sound events, we propose a two-step attention pooling mechanism that provides a time-frequency localisation of multiple audio events in the clip. The proposed framework with two-step attention outperforms existing benchmark models by 22.3%, 12.8%, 5.9% on 0, 10 and 20 dB SNR respectively. We carry out an ablation study to determine the contribution of the auxiliary task and two-step attention pooling to the SED performance improvement.
Although teacher forcing has become the main training paradigm for neural machine translation, it usually makes predictions only conditioned on past information, and hence lacks global planning for the future. To address this problem, we introduce another decoder, called seer decoder, into the encoder-decoder framework during training, which involves future information in target predictions. Meanwhile, we force the conventional decoder to simulate the behaviors of the seer decoder via knowledge distillation. In this way, at test the conventional decoder can perform like the seer decoder without the attendance of it. Experiment results on the Chinese-English, English-German and English-Romanian translation tasks show our method can outperform competitive baselines significantly and achieves greater improvements on the bigger data sets. Besides, the experiments also prove knowledge distillation the best way to transfer knowledge from the seer decoder to the conventional decoder compared to adversarial learning and L2 regularization.
The applications of recurrent neural networks in machine translation are increasing in natural language processing. Besides other languages, Bangla language contains a large amount of vocabulary. Improvement of English to Bangla machine translation would be a significant contribution to Bangla Language processing. This paper describes an architecture of English to Bangla machine translation system. The system has been implemented with the encoder-decoder recurrent neural network. The model uses a knowledge-based context vector for the mapping of English and Bangla words. Performances of the model based on activation functions are measured here. The best performance is achieved for the linear activation function in encoder layer and the tanh activation function in decoder layer. From the execution of GRU and LSTM layer, GRU performed better than LSTM. The attention layers are enacted with softmax and sigmoid activation function. The approach of the model outperforms the previous state-of-the-art systems in terms of cross-entropy loss metrics. The reader can easily find out the structure of the machine translation of English to Bangla and the efficient activation functions from the paper.
Automatic medical image segmentation has made great progress benefit from the development of deep learning. However, most existing methods are based on convolutional neural networks (CNNs), which fail to build long-range dependencies and global context connections due to the limitation of receptive field in convolution operation. Inspired by the success of Transformer in modeling the long-range contextual information, some researchers have expended considerable efforts in designing the robust variants of Transformer-based U-Net. Moreover, the patch division used in vision transformers usually ignores the pixel-level intrinsic structural features inside each patch. To alleviate these problems, we propose a novel deep medical image segmentation framework called Dual Swin Transformer U-Net (DS-TransUNet), which might be the first attempt to concurrently incorporate the advantages of hierarchical Swin Transformer into both encoder and decoder of the standard U-shaped architecture to enhance the semantic segmentation quality of varying medical images. Unlike many prior Transformer-based solutions, the proposed DS-TransUNet first adopts dual-scale encoder subnetworks based on Swin Transformer to extract the coarse and fine-grained feature representations of different semantic scales. As the core component for our DS-TransUNet, a well-designed Transformer Interactive Fusion (TIF) module is proposed to effectively establish global dependencies between features of different scales through the self-attention mechanism. Furthermore, we also introduce the Swin Transformer block into decoder to further explore the long-range contextual information during the up-sampling process. Extensive experiments across four typical tasks for medical image segmentation demonstrate the effectiveness of DS-TransUNet, and show that our approach significantly outperforms the state-of-the-art methods.
Attention-based encoder-decoder framework is widely used in the scene text recognition task. However, for the current state-of-the-art(SOTA) methods, there is room for improvement in terms of the efficient usage of local visual and global context information of the input text image, as well as the robust correlation between the scene processing module(encoder) and the text processing module(decoder). In this paper, we propose a Representation and Correlation Enhanced Encoder-Decoder Framework(RCEED) to address these deficiencies and break performance bottleneck. In the encoder module, local visual feature, global context feature, and position information are aligned and fused to generate a small-size comprehensive feature map. In the decoder module, two methods are utilized to enhance the correlation between scene and text feature space. 1) The decoder initialization is guided by the holistic feature and global glimpse vector exported from the encoder. 2) The feature enriched glimpse vector produced by the Multi-Head General Attention is used to assist the RNN iteration and the character prediction at each time step. Meanwhile, we also design a Layernorm-Dropout LSTM cell to improve model's generalization towards changeable texts. Extensive experiments on the benchmarks demonstrate the advantageous performance of RCEED in scene text recognition tasks, especially the irregular ones.
Medical report generation is one of the most challenging tasks in medical image analysis. Although existing approaches have achieved promising results, they either require a predefined template database in order to retrieve sentences or ignore the hierarchical nature of medical report generation. To address these issues, we propose MedWriter that incorporates a novel hierarchical retrieval mechanism to automatically extract both report and sentence-level templates for clinically accurate report generation. MedWriter first employs the Visual-Language Retrieval~(VLR) module to retrieve the most relevant reports for the given images. To guarantee the logical coherence between sentences, the Language-Language Retrieval~(LLR) module is introduced to retrieve relevant sentences based on the previous generated description. At last, a language decoder fuses image features and features from retrieved reports and sentences to generate meaningful medical reports. We verified the effectiveness of our model by automatic evaluation and human evaluation on two datasets, i.e., Open-I and MIMIC-CXR.
Deep learning has become the dominant approach in coping with various tasks in Natural LanguageProcessing (NLP). Although text inputs are typically represented as a sequence of tokens, there isa rich variety of NLP problems that can be best expressed with a graph structure. As a result, thereis a surge of interests in developing new deep learning techniques on graphs for a large numberof NLP tasks. In this survey, we present a comprehensive overview onGraph Neural Networks(GNNs) for Natural Language Processing. We propose a new taxonomy of GNNs for NLP, whichsystematically organizes existing research of GNNs for NLP along three axes: graph construction,graph representation learning, and graph based encoder-decoder models. We further introducea large number of NLP applications that are exploiting the power of GNNs and summarize thecorresponding benchmark datasets, evaluation metrics, and open-source codes. Finally, we discussvarious outstanding challenges for making the full use of GNNs for NLP as well as future researchdirections. To the best of our knowledge, this is the first comprehensive overview of Graph NeuralNetworks for Natural Language Processing.
This paper describes a novel lossless point cloud compression algorithm that uses a neural network for estimating the coding probabilities for the occupancy status of voxels, depending on wide three dimensional contexts around the voxel to be encoded. The point cloud is represented as an octree, with each resolution layer being sequentially encoded and decoded using arithmetic coding, starting from the lowest resolution, until the final resolution is reached. The occupancy probability of each voxel of the splitting pattern at each node of the octree is modeled by a neural network, having at its input the already encoded occupancy status of several octree nodes (belonging to the past and current resolutions), corresponding to a 3D context surrounding the node to be encoded. The algorithm has a fast and a slow version, the fast version selecting differently several voxels of the context, which allows an increased parallelization by sending larger batches of templates to be estimated by the neural network, at both encoder and decoder. The proposed algorithms yield state-of-the-art results on benchmark datasets. The implementation will be made available at https://github.com/marmus12/nnctx