Piece-wise 3D planar reconstruction provides holistic scene understanding of man-made environments, especially for indoor scenarios. Most recent approaches focused on improving the segmentation and reconstruction results by introducing advanced network architectures but overlooked the dual characteristics of piece-wise planes as objects and geometric models. Different from other existing approaches, we start from enforcing cross-task consistency for our multi-task convolutional neural network, PlaneRecNet, which integrates a single-stage instance segmentation network for piece-wise planar segmentation and a depth decoder to reconstruct the scene from a single RGB image. To achieve this, we introduce several novel loss functions (geometric constraint) that jointly improve the accuracy of piece-wise planar segmentation and depth estimation. Meanwhile, a novel Plane Prior Attention module is used to guide depth estimation with the awareness of plane instances. Exhaustive experiments are conducted in this work to validate the effectiveness and efficiency of our method.
Currently, existing salient object detection methods based on convolutional neural networks commonly resort to constructing discriminative networks to aggregate high level and low level features. However, contextual information is always not fully and reasonably utilized, which usually causes either the absence of useful features or contamination of redundant features. To address these issues, we propose a novel ladder context correlation complementary network (LC3Net) in this paper, which is equipped with three crucial components. At the beginning, we propose a filterable convolution block (FCB) to assist the automatic collection of information on the diversity of initial features, and it is simple yet practical. Besides, we propose a dense cross module (DCM) to facilitate the intimate aggregation of different levels of features by validly integrating semantic information and detailed information of both adjacent and non-adjacent layers. Furthermore, we propose a bidirectional compression decoder (BCD) to help the progressive shrinkage of multi-scale features from coarse to fine by leveraging multiple pairs of alternating top-down and bottom-up feature interaction flows. Extensive experiments demonstrate the superiority of our method against 16 state-of-the-art methods.
Explainable recommendation systems provide explanations for recommendation results to improve their transparency and persuasiveness. The existing explainable recommendation methods generate textual explanations without explicitly considering the user's preferences on different aspects of the item. In this paper, we propose a novel explanation generation framework, named Hierarchical Aspect-guided explanation Generation (HAG), for explainable recommendation. Specifically, HAG employs a review-based syntax graph to provide a unified view of the user/item details. An aspect-guided graph pooling operator is proposed to extract the aspect-relevant information from the review-based syntax graphs to model the user's preferences on an item at the aspect level. Then, a hierarchical explanation decoder is developed to generate aspects and aspect-relevant explanations based on the attention mechanism. The experimental results on three real datasets indicate that HAG outperforms state-of-the-art explanation generation methods in both single-aspect and multi-aspect explanation generation tasks, and also achieves comparable or even better preference prediction accuracy than strong baseline methods.
The application of deep learning in survival analysis (SA) allows utilizing unstructured and high-dimensional data types uncommon in traditional survival methods. This allows to advance methods in fields such as digital health, predictive maintenance, and churn analysis, but often yields less interpretable and intuitively understandable models due to the black-box character of deep learning-based approaches. We close this gap by proposing 1) a multi-task variational autoencoder (VAE) with survival objective, yielding survival-oriented embeddings, and 2) a novel method HazardWalk that allows to model hazard factors in the original data space. HazardWalk transforms the latent distribution of our autoencoder into areas of maximized/minimized hazard and then uses the decoder to project changes to the original domain. Our procedure is evaluated on a simulated dataset as well as on a dataset of CT imaging data of patients with liver metastases.
Accurate and rapid detection of COVID-19 pneumonia is crucial for optimal patient treatment. Chest X-Ray (CXR) is the first line imaging test for COVID-19 pneumonia diagnosis as it is fast, cheap and easily accessible. Inspired by the success of deep learning (DL) in computer vision, many DL-models have been proposed to detect COVID-19 pneumonia using CXR images. Unfortunately, these deep classifiers lack the transparency in interpreting findings, which may limit their applications in clinical practice. The existing commonly used visual explanation methods are either too noisy or imprecise, with low resolution, and hence are unsuitable for diagnostic purposes. In this work, we propose a novel explainable deep learning framework (CXRNet) for accurate COVID-19 pneumonia detection with an enhanced pixel-level visual explanation from CXR images. The proposed framework is based on a new Encoder-Decoder-Encoder multitask architecture, allowing for both disease classification and visual explanation. The method has been evaluated on real world CXR datasets from both public and private data sources, including: healthy, bacterial pneumonia, viral pneumonia and COVID-19 pneumonia cases The experimental results demonstrate that the proposed method can achieve a satisfactory level of accuracy and provide fine-resolution classification activation maps for visual explanation in lung disease detection. The Average Accuracy, the Precision, Recall and F1-score of COVID-19 pneumonia reached 0.879, 0.985, 0.992 and 0.989, respectively. We have also found that using lung segmented (CXR) images can help improve the performance of the model. The proposed method can provide more detailed high resolution visual explanation for the classification decision, compared to current state-of-the-art visual explanation methods and has a great potential to be used in clinical practice for COVID-19 pneumonia diagnosis.
Gaussian processes (GPs), implemented through multivariate Gaussian distributions for a finite collection of data, are the most popular approach in small-area spatial statistical modelling. In this context they are used to encode correlation structures over space and can generalise well in interpolation tasks. Despite their flexibility, off-the-shelf GPs present serious computational challenges which limit their scalability and practical usefulness in applied settings. Here, we propose a novel, deep generative modelling approach to tackle this challenge, termed PriorVAE: for a particular spatial setting, we approximate a class of GP priors through prior sampling and subsequent fitting of a variational autoencoder (VAE). Given a trained VAE, the resultant decoder allows spatial inference to become incredibly efficient due to the low dimensional, independently distributed latent Gaussian space representation of the VAE. Once trained, inference using the VAE decoder replaces the GP within a Bayesian sampling framework. This approach provides tractable and easy-to-implement means of approximately encoding spatial priors and facilitates efficient statistical inference. We demonstrate the utility of our VAE two stage approach on Bayesian, small-area estimation tasks.
Recent advances in transformer-based models have drawn attention to exploring these techniques in medical image segmentation, especially in conjunction with the U-Net model (or its variants), which has shown great success in medical image segmentation, under both 2D and 3D settings. Current 2D based methods either directly replace convolutional layers with pure transformers or consider a transformer as an additional intermediate encoder between the encoder and decoder of U-Net. However, these approaches only consider the attention encoding within one single slice and do not utilize the axial-axis information naturally provided by a 3D volume. In the 3D setting, convolution on volumetric data and transformers both consume large GPU memory. One has to either downsample the image or use cropped local patches to reduce GPU memory usage, which limits its performance. In this paper, we propose Axial Fusion Transformer UNet (AFTer-UNet), which takes both advantages of convolutional layers' capability of extracting detailed features and transformers' strength on long sequence modeling. It considers both intra-slice and inter-slice long-range cues to guide the segmentation. Meanwhile, it has fewer parameters and takes less GPU memory to train than the previous transformer-based models. Extensive experiments on three multi-organ segmentation datasets demonstrate that our method outperforms current state-of-the-art methods.
Existing text- and image-based multimodal dialogue systems use the traditional Hierarchical Recurrent Encoder-Decoder (HRED) framework, which has an utterance-level encoder to model utterance representation and a context-level encoder to model context representation. Although pioneer efforts have shown promising performances, they still suffer from the following challenges: (1) the interaction between textual features and visual features is not fine-grained enough. (2) the context representation can not provide a complete representation for the context. To address the issues mentioned above, we propose a non-hierarchical attention network with modality dropout, which abandons the HRED framework and utilizes attention modules to encode each utterance and model the context representation. To evaluate our proposed model, we conduct comprehensive experiments on a public multimodal dialogue dataset. Automatic and human evaluation demonstrate that our proposed model outperforms the existing methods and achieves state-of-the-art performance.
Polar codes are high density parity check codes and hence the sparse factor graph, instead of the parity check matrix, has been used to practically represent an LP polytope for LP decoding. Although LP decoding on this polytope has the ML-certificate property, it performs poorly over a BAWGN channel. In this paper, we propose modifications to adaptive cut generation based LP decoding techniques and apply the modified- adaptive LP decoder to short blocklength polar codes over a BAWGN channel. The proposed decoder provides significant FER performance gain compared to the previously proposed LP decoder and its performance approaches that of ML decoding at high SNRs. We also present an algorithm to obtain a smaller factor graph from the original sparse factor graph of a polar code. This reduced factor graph preserves the small check node degrees needed to represent the LP polytope in practice. We show that the fundamental polytope of the reduced factor graph can be obtained from the projection of the polytope represented by the original sparse factor graph and the frozen bit information. Thus, the LP decoding time complexity is decreased without changing the FER performance by using the reduced factor graph representation.
In this paper, we propose VISinger, a complete end-to-end high-quality singing voice synthesis (SVS) system that directly generates audio waveform from lyrics and musical score. Our approach is inspired by VITS, which adopts VAE-based posterior encoder augmented with normalizing flow-based prior encoder and adversarial decoder to realize complete end-to-end speech generation. VISinger follows the main architecture of VITS, but makes substantial improvements to the prior encoder based on the characteristics of singing. First, instead of using phoneme-level mean and variance of acoustic features, we introduce a length regulator and a frame prior network to get the frame-level mean and variance on acoustic features, modeling the rich acoustic variation in singing. Second, we further introduce an F0 predictor to guide the frame prior network, leading to stabler singing performance. Finally, to improve the singing rhythm, we modify the duration predictor to specifically predict the phoneme to note duration ratio, helped with singing note normalization. Experiments on a professional Mandarin singing corpus show that VISinger significantly outperforms FastSpeech+Neural-Vocoder two-stage approach and the oracle VITS; ablation study demonstrates the effectiveness of different contributions.
Time series data is ubiquitous in research as well as in a wide variety of industrial applications. Effectively analyzing the available historical data and providing insights into the far future allows us to make effective decisions. Recent research has witnessed the superior performance of transformer-based architectures, especially in the regime of far horizon time series forecasting. However, the current state of the art sparse Transformer architectures fail to couple down- and upsampling procedures to produce outputs in a similar resolution as the input. We propose the Yformer model, based on a novel Y-shaped encoder-decoder architecture that (1) uses direct connection from the downscaled encoder layer to the corresponding upsampled decoder layer in a U-Net inspired architecture, (2) Combines the downscaling/upsampling with sparse attention to capture long-range effects, and (3) stabilizes the encoder-decoder stacks with the addition of an auxiliary reconstruction loss. Extensive experiments have been conducted with relevant baselines on four benchmark datasets, demonstrating an average improvement of 19.82, 18.41 percentage MSE and 13.62, 11.85 percentage MAE in comparison to the current state of the art for the univariate and the multivariate settings respectively.
To improve the performance of state-of-the-art automatic speech recognition systems it is common practice to include external knowledge sources such as language models or prior corrections. This is usually done via log-linear model combination using separate scaling parameters for each model. Typically these parameters are manually optimized on some held-out data. In this work we propose to optimize these scaling parameters via automatic differentiation and stochastic gradient decent similar to the neural network model parameters. We show on the LibriSpeech (LBS) and Switchboard (SWB) corpora that the model scales for a combination of attentionbased encoder-decoder acoustic model and language model can be learned as effectively as with manual tuning. We further extend this approach to subword dependent model scales which could not be tuned manually which leads to 7% improvement on LBS and 3% on SWB. We also show that joint training of scales and model parameters is possible and gives additional 6% improvement on LBS.
Building Spoken Language Understanding (SLU) systems that do not rely on language specific Automatic Speech Recognition (ASR) is an important yet less explored problem in language processing. In this paper, we present a comparative study aimed at employing a pre-trained acoustic model to perform SLU in low resource scenarios. Specifically, we use three different embeddings extracted using Allosaurus, a pre-trained universal phone decoder: (1) Phone (2) Panphone, and (3) Allo embeddings. These embeddings are then used in identifying the spoken intent. We perform experiments across three different languages: English, Sinhala, and Tamil each with different data sizes to simulate high, medium, and low resource scenarios. Our system improves on the state-of-the-art (SOTA) intent classification accuracy by approximately 2.11% for Sinhala and 7.00% for Tamil and achieves competitive results on English. Furthermore, we present a quantitative analysis of how the performance scales with the number of training examples used per intent.
Sequence discriminative training is a great tool to improve the performance of an automatic speech recognition system. It does, however, necessitate a sum over all possible word sequences, which is intractable to compute in practice. Current state-of-the-art systems with unlimited label context circumvent this problem by limiting the summation to an n-best list of relevant competing hypotheses obtained from beam search. This work proposes to perform (approximative) recombinations of hypotheses during beam search, if they share a common local history. The error that is incurred by the approximation is analyzed and it is shown that using this technique the effective beam size can be increased by several orders of magnitude without significantly increasing the computational requirements. Lastly, it is shown that this technique can be used to effectively perform sequence discriminative training for attention-based encoder-decoder acoustic models on the LibriSpeech task.
This paper demonstrates that multilingual pretraining and multilingual fine-tuning are both critical for facilitating cross-lingual transfer in zero-shot translation, where the neural machine translation (NMT) model is tested on source languages unseen during supervised training. Following this idea, we present SixT+, a strong many-to-English NMT model that supports 100 source languages but is trained with a parallel dataset in only six source languages. SixT+ initializes the decoder embedding and the full encoder with XLM-R large and then trains the encoder and decoder layers with a simple two-stage training strategy. SixT+ achieves impressive performance on many-to-English translation. It significantly outperforms CRISS and m2m-100, two strong multilingual NMT systems, with an average gain of 7.2 and 5.0 BLEU respectively. Additionally, SixT+ offers a set of model parameters that can be further fine-tuned to other unsupervised tasks. We demonstrate that adding SixT+ initialization outperforms state-of-the-art explicitly designed unsupervised NMT models on Si<->En and Ne<->En by over 1.2 average BLEU. When applied to zero-shot cross-lingual abstractive summarization, it produces an average performance gain of 12.3 ROUGE-L over mBART-ft. We conduct detailed analyses to understand the key ingredients of SixT+, including multilinguality of the auxiliary parallel data, positional disentangled encoder, and the cross-lingual transferability of its encoder.
We introduce PRIMERA, a pre-trained model for multi-document representation with a focus on summarization that reduces the need for dataset-specific architectures and large amounts of fine-tuning labeled data. PRIMERA uses our newly proposed pre-training objective designed to teach the model to connect and aggregate information across documents. It also uses efficient encoder-decoder transformers to simplify the processing of concatenated input documents. With extensive experiments on 6 multi-document summarization datasets from 3 different domains on zero-shot, few-shot and full-supervised settings, PRIMERA outperforms current state-of-the-art dataset-specific and pre-trained models on most of these settings with large margins. The code and pre-trained models can be found at \url{https://github.com/allenai/PRIMER}.
Pre-trained Language Models (PLMs) have been successful for a wide range of natural language processing (NLP) tasks. The state-of-the-art of PLMs, however, are extremely large to be used on edge devices. As a result, the topic of model compression has attracted increasing attention in the NLP community. Most of the existing works focus on compressing encoder-based models (tiny-BERT, distilBERT, distilRoBERTa, etc), however, to the best of our knowledge, the compression of decoder-based models (such as GPT-2) has not been investigated much. Our paper aims to fill this gap. Specifically, we explore two directions: 1) we employ current state-of-the-art knowledge distillation techniques to improve fine-tuning of DistilGPT-2. 2) we pre-train a compressed GPT-2 model using layer truncation and compare it against the distillation-based method (DistilGPT2). The training time of our compressed model is significantly less than DistilGPT-2, but it can achieve better performance when fine-tuned on downstream tasks. We also demonstrate the impact of data cleaning on model performance.
Pre-trained encoder-decoder transformer architectures have become increasingly popular recently with the advent of T5 models. T5 has also become more favorable over other architectures like BERT due to the amount of data that it is pre-trained on, increased scale of model parameter sizes and easy applicability to a diverse set of tasks due to the generative nature of the model. While being able to generalize to a wide variety of tasks, it is not clear that encoder-decoder architectures are the most efficient for fine-tuning tasks that don't require auto-regressive decoding. In this work, we study fine-tuning pre-trained encoder-decoder models for tasks such as classification, multi-label classification, and structured prediction. We propose \textbf{EncT5}, a framework for these problems, and illustrate instantiations for these tasks. Our experiment results show that EncT5 has advantages over T5 such as efficiency and usability out performs BERT when evaluated on publicly available pre-trained checkpoints.
Temporally consistent dense video annotations are scarce and hard to collect. In contrast, image segmentation datasets (and pre-trained models) are ubiquitous, and easier to label for any novel task. In this paper, we introduce a method to adapt still image segmentation models to video in an unsupervised manner, by using an optical flow-based consistency measure. To ensure that the inferred segmented videos appear more stable in practice, we verify that the consistency measure is well correlated with human judgement via a user study. Training a new multi-input multi-output decoder using this measure as a loss, together with a technique for refining current image segmentation datasets and a temporal weighted-guided filter, we observe stability improvements in the generated segmented videos with minimal loss of accuracy.
3D reconstruction aims to reconstruct 3D objects from 2D views. Previous works for 3D reconstruction mainly focus on feature matching between views or using CNNs as backbones. Recently, Transformers have been shown effective in multiple applications of computer vision. However, whether or not Transformers can be used for 3D reconstruction is still unclear. In this paper, we fill this gap by proposing 3D-RETR, which is able to perform end-to-end 3D REconstruction with TRansformers. 3D-RETR first uses a pretrained Transformer to extract visual features from 2D input images. 3D-RETR then uses another Transformer Decoder to obtain the voxel features. A CNN Decoder then takes as input the voxel features to obtain the reconstructed objects. 3D-RETR is capable of 3D reconstruction from a single view or multiple views. Experimental results on two datasets show that 3DRETR reaches state-of-the-art performance on 3D reconstruction. Additional ablation study also demonstrates that 3D-DETR benefits from using Transformers.