Does a Variational AutoEncoder (VAE) consistently encode typical samples generated from its decoder? This paper shows that the perhaps surprising answer to this question is `No'; a (nominally trained) VAE does not necessarily amortize inference for typical samples that it is capable of generating. We study the implications of this behaviour on the learned representations and also the consequences of fixing it by introducing a notion of self consistency. Our approach hinges on an alternative construction of the variational approximation distribution to the true posterior of an extended VAE model with a Markov chain alternating between the encoder and the decoder. The method can be used to train a VAE model from scratch or given an already trained VAE, it can be run as a post processing step in an entirely self supervised way without access to the original training data. Our experimental analysis reveals that encoders trained with our self-consistency approach lead to representations that are robust (insensitive) to perturbations in the input introduced by adversarial attacks. We provide experimental results on the ColorMnist and CelebA benchmark datasets that quantify the properties of the learned representations and compare the approach with a baseline that is specifically trained for the desired property.
Unconstrained handwritten text recognition remains challenging for computer vision systems. Paragraph text recognition is traditionally achieved by two models: the first one for line segmentation and the second one for text line recognition. We propose a unified end-to-end model using hybrid attention to tackle this task. This model is designed to iteratively process a paragraph image line by line. It can be split into three modules. An encoder generates feature maps from the whole paragraph image. Then, an attention module recurrently generates a vertical weighted mask enabling to focus on the current text line features. This way, it performs a kind of implicit line segmentation. For each text line features, a decoder module recognizes the character sequence associated, leading to the recognition of a whole paragraph. We achieve state-of-the-art character error rate at paragraph level on three popular datasets: 1.91% for RIMES, 4.45% for IAM and 3.59% for READ 2016. Our code and trained model weights are available at https://github.com/FactoDeepLearning/VerticalAttentionOCR.
Automatic segmentation of brain glioma from multimodal MRI scans plays a key role in clinical trials and practice. Unfortunately, manual segmentation is very challenging, time-consuming, costly, and often inaccurate despite human expertise due to the high variance and high uncertainty in the human annotations. In the present work, we develop an end-to-end deep-learning-based segmentation method using a multi-decoder architecture by jointly learning three separate sub-problems using a partly shared encoder. We also propose to apply smoothing methods to the input images to generate denoised versions as additional inputs to the network. The validation performance indicate an improvement when using the proposed method. The proposed method was ranked 2nd in the task of Quantification of Uncertainty in Segmentation in the Brain Tumors in Multimodal Magnetic Resonance Imaging Challenge 2020.
The interpretation of seismic data is vital for characterizing sediments' shape in areas of geological study. In seismic interpretation, deep learning becomes useful for reducing the dependence on handcrafted facies segmentation geometry and the time required to study geological areas. This work presents a Deep Neural Network for Facies Segmentation (DNFS) to obtain state-of-the-art results for seismic facies segmentation. DNFS is trained using a combination of cross-entropy and Jaccard loss functions. Our results show that DNFS obtains highly detailed predictions for seismic facies segmentation using fewer parameters than StNet and U-Net.
Because the expansion path of U-Net may ignore the characteristics of small targets, intermediate supervision mechanism is proposed. The original mask is also entered into the network as a label for intermediate output. However, U-Net is mainly engaged in segmentation, and the extracted features are also targeted at segmentation location information, and the input and output are different. The label we need is that the input and output are both original masks, which is more similar to the refactoring process, so we propose another intermediate supervision mechanism. However, the features extracted by the contraction path of this intermediate monitoring mechanism are not necessarily consistent. For example, U-Net's contraction path extracts transverse features, while auto-encoder extracts longitudinal features, which may cause the output of the expansion path to be inconsistent with the label. Therefore, we put forward the intermediate supervision mechanism of shared-weight decoder module. Although the intermediate supervision mechanism improves the segmentation accuracy, the training time is too long due to the extra input and multiple loss functions. For one of these problems, we have introduced tied-weight decoder. To reduce the redundancy of the model, we combine shared-weight decoder module with tied-weight decoder module.
This paper introduces a graphical representation approach of prosody boundary (GraphPB) in the task of Chinese speech synthesis, intending to parse the semantic and syntactic relationship of input sequences in a graphical domain for improving the prosody performance. The nodes of the graph embedding are formed by prosodic words, and the edges are formed by the other prosodic boundaries, namely prosodic phrase boundary (PPH) and intonation phrase boundary (IPH). Different Graph Neural Networks (GNN) like Gated Graph Neural Network (GGNN) and Graph Long Short-term Memory (G-LSTM) are utilised as graph encoders to exploit the graphical prosody boundary information. Graph-to-sequence model is proposed and formed by a graph encoder and an attentional decoder. Two techniques are proposed to embed sequential information into the graph-to-sequence text-to-speech model. The experimental results show that this proposed approach can encode the phonetic and prosody rhythm of an utterance. The mean opinion score (MOS) of these GNN models shows comparative results with the state-of-the-art sequence-to-sequence models with better performance in the aspect of prosody. This provides an alternative approach for prosody modelling in end-to-end speech synthesis.
We propose a novel general method that finds action-grounded, discrete object and effect categories and builds probabilistic rules over them for non-trivial action planning. Our robot interacts with objects using an initial action repertoire that is assumed to be acquired earlier and observes the effects it can create in the environment. To form action-grounded object, effect, and relational categories, we employ a binary bottleneck layer in a predictive, deep encoder-decoder network that takes the image of the scene and the action applied as input, and generates the resulting effects in the scene in pixel coordinates. After learning, the binary latent vector represents action-driven object categories based on the interaction experience of the robot. To distill the knowledge represented by the neural network into rules useful for symbolic reasoning, a decision tree is trained to reproduce its decoder function. Probabilistic rules are extracted from the decision paths of the tree and are represented in the Probabilistic Planning Domain Definition Language (PPDDL), allowing off-the-shelf planners to operate on the knowledge extracted from the sensorimotor experience of the robot. The deployment of the proposed approach for a simulated robotic manipulator enabled the discovery of discrete representations of object properties such as `rollable' and `insertable'. In turn, the use of these representations as symbols allowed the generation of effective plans for achieving goals, such as building towers of the desired height, demonstrating the effectiveness of the approach for multi-step object manipulation. Finally, we demonstrate that the system is not only restricted to the robotics domain by assessing its applicability to the MNIST 8-puzzle domain in which learned symbols allow for the generation of plans that move the empty tile into any given position.
Can AI help automate human-easy but computer-hard data preparation tasks that burden data scientists, practitioners, and crowd workers? We answer this question by presenting RPT, a denoising auto-encoder for tuple-to-X models (X could be tuple, token, label, JSON, and so on). RPT is pre-trained for a tuple-to-tuple model by corrupting the input tuple and then learning a model to reconstruct the original tuple. It adopts a Transformer-based neural translation architecture that consists of a bidirectional encoder (similar to BERT) and a left-to-right autoregressive decoder (similar to GPT), leading to a generalization of both BERT and GPT. The pre-trained RPT can already support several common data preparation tasks such as data cleaning, auto-completion and schema matching. Better still, RPT can be fine-tuned on a wide range of data preparation tasks, such as value normalization, data transformation, data annotation, etc. To complement RPT, we also discuss several appealing techniques such as collaborative training and few-shot learning for entity resolution, and few-shot learning and NLP question-answering for information extraction. In addition, we identify a series of research opportunities to advance the field of data preparation.
Next destination recommendation is an important task in the transportation domain of taxi and ride-hailing services, where users are recommended with personalized destinations given their current origin location. However, recent recommendation works do not satisfy this origin-awareness property, and only consider learning from historical destination locations, without origin information. Thus, the resulting approaches are unable to learn and predict origin-aware recommendations based on the user's current location, leading to sub-optimal performance and poor real-world practicality. Hence, in this work, we study the origin-aware next destination recommendation task. We propose the Spatial-Temporal Origin-Destination Personalized Preference Attention (STOD-PPA) encoder-decoder model to learn origin-origin (OO), destination-destination (DD), and origin-destination (OD) relationships by first encoding both origin and destination sequences with spatial and temporal factors in local and global views, then decoding them through personalized preference attention to predict the next destination. Experimental results on seven real-world user trajectory taxi datasets show that our model significantly outperforms baseline and state-of-the-art methods.
Automatic lung lesions segmentation of chest CT scans is considered a pivotal stage towards accurate diagnosis and severity measurement of COVID-19. Traditional U-shaped encoder-decoder architecture and its variants suffer from diminutions of contextual information in pooling/upsampling operations with increased semantic gaps among encoded and decoded feature maps as well as instigate vanishing gradient problems for its sequential gradient propagation that result in sub-optimal performance. Moreover, operating with 3D CT-volume poses further limitations due to the exponential increase of computational complexity making the optimization difficult. In this paper, an automated COVID-19 lesion segmentation scheme is proposed utilizing a highly efficient neural network architecture, namely CovSegNet, to overcome these limitations. Additionally, a two-phase training scheme is introduced where a deeper 2D-network is employed for generating ROI-enhanced CT-volume followed by a shallower 3D-network for further enhancement with more contextual information without increasing computational burden. Along with the traditional vertical expansion of Unet, we have introduced horizontal expansion with multi-stage encoder-decoder modules for achieving optimum performance. Additionally, multi-scale feature maps are integrated into the scale transition process to overcome the loss of contextual information. Moreover, a multi-scale fusion module is introduced with a pyramid fusion scheme to reduce the semantic gaps between subsequent encoder/decoder modules while facilitating the parallel optimization for efficient gradient propagation. Outstanding performances have been achieved in three publicly available datasets that largely outperform other state-of-the-art approaches. The proposed scheme can be easily extended for achieving optimum segmentation performances in a wide variety of applications.
Current state-of-the-art methods for skeleton-based action recognition are supervised and rely on labels. The reliance is limiting the performance due to the challenges involved in annotation and mislabeled data. Unsupervised methods have been introduced, however, they organize sequences into clusters and still require labels to associate clusters with actions. In this paper, we propose a novel approach for skeleton-based action recognition, called SESAR, that connects these approaches. SESAR leverages the information from both unlabeled data and a handful of sequences actively selected for labeling, combining unsupervised training with sparsely supervised guidance. SESAR is composed of two main components, where the first component learns a latent representation for unlabeled action sequences through an Encoder-Decoder RNN which reconstructs the sequences, and the second component performs active learning to select sequences to be labeled based on cluster and classification uncertainty. When the two components are simultaneously trained on skeleton-based action sequences, they correspond to a robust system for action recognition with only a handful of labeled samples. We evaluate our system on common datasets with multiple sequences and actions, such as NW UCLA, NTU RGB+D 60, and UWA3D. Our results outperform standalone skeleton-based supervised, unsupervised with cluster identification, and active-learning methods for action recognition when applied to sparse labeled samples, as low as 1% of the data.
We consider the task of representation learning for unsupervised segmentation of 3D voxel-grid biomedical images. We show that models that capture implicit hierarchical relationships between subvolumes are better suited for this task. To that end, we consider encoder-decoder architectures with a hyperbolic latent space, to explicitly capture hierarchical relationships present in subvolumes of the data. We propose utilizing a 3D hyperbolic variational autoencoder with a novel gyroplane convolutional layer to map from the embedding space back to 3D images. To capture these relationships, we introduce an essential self-supervised loss -- in addition to the standard VAE loss -- which infers approximate hierarchies and encourages implicitly related subvolumes to be mapped closer in the embedding space. We present experiments on both synthetic data and biomedical data to validate our hypothesis.
We propose an online multi-view depth prediction approach on posed video streams, where the scene geometry information computed in the previous time steps is propagated to the current time step in an efficient and geometrically plausible way. The backbone of our approach is a real-time capable, lightweight encoder-decoder that relies on cost volumes computed from pairs of images. We extend it by placing a ConvLSTM cell at the bottleneck layer, which compresses an arbitrary amount of past information in its states. The novelty lies in propagating the hidden state of the cell by accounting for the viewpoint changes between time steps. At a given time step, we warp the previous hidden state into the current camera plane using the previous depth prediction. Our extension brings only a small overhead of computation time and memory consumption, while improving the depth predictions significantly. As a result, we outperform the existing state-of-the-art multi-view stereo methods on most of the evaluated metrics in hundreds of indoor scenes while maintaining a real-time performance. Code available: https://github.com/ardaduz/deep-video-mvs
Machine learning on trees has been mostly focused on trees as input to algorithms. Much less research has investigated trees as output, which has many applications, such as molecule optimization for drug discovery, or hint generation for intelligent tutoring systems. In this work, we propose a novel autoencoder approach, called recursive tree grammar autoencoder (RTG-AE), which encodes trees via a bottom-up parser and decodes trees via a tree grammar, both learned via recursive neural networks that minimize the variational autoencoder loss. The resulting encoder and decoder can then be utilized in subsequent tasks, such as optimization and time series prediction. RTG-AEs are the first model to combine variational autoencoders, grammatical knowledge, and recursive processing. Our key message is that this unique combination of all three elements outperforms models which combine any two of the three. In particular, we perform an ablation study to show that our proposed method improves the autoencoding error, training time, and optimization score on synthetic as well as real datasets compared to four baselines. We further prove that RTG-AEs parse and generate trees in linear time and are expressive enough to handle all regular tree grammars.
Abstractive summarization is the process of generating novel sentences based on the information extracted from the original text document while retaining the context. Due to abstractive summarization's underlying complexities, most of the past research work has been done on the extractive summarization approach. Nevertheless, with the triumph of the sequence-to-sequence (seq2seq) model, abstractive summarization becomes more viable. Although a significant number of notable research has been done in the English language based on abstractive summarization, only a couple of works have been done on Bengali abstractive news summarization (BANS). In this article, we presented a seq2seq based Long Short-Term Memory (LSTM) network model with attention at encoder-decoder. Our proposed system deploys a local attention-based model that produces a long sequence of words with lucid and human-like generated sentences with noteworthy information of the original document. We also prepared a dataset of more than 19k articles and corresponding human-written summaries collected from bangla.bdnews24.com1 which is till now the most extensive dataset for Bengali news document summarization and publicly published in Kaggle2. We evaluated our model qualitatively and quantitatively and compared it with other published results. It showed significant improvement in terms of human evaluation scores with state-of-the-art approaches for BANS.
Stimulation methods that utilise more than one stimulation frequency have been developed for steady-state visual evoked potential (SSVEP) brain-computer interfaces (BCIs) with the purpose of increasing the number of targets that can be presented simultaneously. However, there is no unified decoding algorithm that can be used without training for each individual users or cases, and applied to a large class of multi-frequency stimulated SSVEP settings. This paper extends the widely used canonical correlation analysis (CCA) decoder to explicitly accommodate multi-frequency SSVEP by exploiting the interactions between the multiple stimulation frequencies. A concept of order, defined as the sum of absolute value of the coefficients in the linear combination of the input frequencies, was introduced to assist the design of Multi-Frequency CCA (MFCCA). The probability distribution of the order in the resulting SSVEP response was then used to improve decoding accuracy. Results show that, compared to the standard CCA formulation, the proposed MFCCA has a 20% improvement in decoding accuracy on average at order 2, while keeping its generality and training-free characteristics.
With the number of smart devices increasing, the demand for on-device text-to-speech (TTS) increases rapidly. In recent years, many prominent End-to-End TTS methods have been proposed, and have greatly improved the quality of synthesized speech. However, to ensure the qualified speech, most TTS systems depend on large and complex neural network models, and it's hard to deploy these TTS systems on-device. In this paper, a small-footprint, fast, stable network for on-device TTS is proposed, named as DeviceTTS. DeviceTTS makes use of a duration predictor as a bridge between encoder and decoder so as to avoid the problem of words skipping and repeating in Tacotron. As we all know, model size is a key factor for on-device TTS. For DeviceTTS, Deep Feedforward Sequential Memory Network (DFSMN) is used as the basic component. Moreover, to speed up inference, mix-resolution decoder is proposed for balance the inference speed and speech quality. Experiences are done with WORLD and LPCNet vocoder. Finally, with only 1.4 million model parameters and 0.099 GFLOPS, DeviceTTS achieves comparable performance with Tacotron and FastSpeech. As far as we know, the DeviceTTS can meet the needs of most of the devices in practical application.
The attention-based encoder-decoder modeling paradigm has achieved promising results on a variety of speech processing tasks like automatic speech recognition (ASR), text-to-speech (TTS) and among others. This paradigm takes advantage of the generalization ability of neural networks to learn a direct mapping from an input sequence to an output sequence, without recourse to prior knowledge such as audio-text alignments or pronunciation lexicons. However, ASR models stemming from this paradigm are prone to overfitting, especially when the training data is limited. Inspired by SpecAugment and BERT-like masked language modeling, we propose in the paper a decoder masking based training approach for end-to-end (E2E) ASR models. During the training phase we randomly replace some portions of the decoder's historical text input with the symbol [mask], in order to encourage the decoder to robustly output a correct token even when parts of its decoding history are masked or corrupted. The proposed approach is instantiated with the top-of-the-line transformer-based E2E ASR model. Extensive experiments on the Librispeech960h and TedLium2 benchmark datasets demonstrate the superior performance of our approach in comparison to some existing strong E2E ASR systems.
This paper presents SSLIDE, Sound Source Localization for Indoors using DEep learning, which applies deep neural networks (DNNs) with encoder-decoder structure to localize sound sources with random positions in a continuous space. The spatial features of sound signals received by each microphone are extracted and represented as likelihood surfaces for the sound source locations in each point. Our DNN consists of an encoder network followed by two decoders. The encoder obtains a compressed representation of the input likelihoods. One decoder resolves the multipath caused by reverberation, and the other decoder estimates the source location. Experiments based on both the simulated and experimental data show that our method can not only outperform multiple signal classification (MUSIC), steered response power with phase transform (SRP-PHAT), sparse Bayesian learning (SBL), and a competing convolutional neural network (CNN) approach in the reverberant environment but also achieve a good generalization performance.
Objective. Brain-computer interfaces (BCIs) create a new communication pathway between the brain and an effector without neuromuscular activation. BCI experiments highlighted high intra and inter-subjects variability in the BCI decoders. Although BCI model is generally relying on neurological markers generalizable on the majority of subjects, it requires to generate a wide range of neural features to include possible neurophysiological patterns. However, the processing of noisy and high dimensional features, such as brain signals, brings several challenges to overcome such as model calibration issues, model generalization and interpretation problems and hardware related obstacles. Approach. An online adaptive group-wise sparse decoder named Lp-Penalized REW-NPLS algorithm (PREW-NPLS) is presented to reduce the feature space dimension employed for BCI decoding. The proposed decoder was designed to create BCI systems with low computational cost suited for portable applications and tested during offline pseudo-online study based on online closed-loop BCI control of the left and right 3D arm movements of a virtual avatar from the ECoG recordings of a tetraplegic patient. Main results. PREW-NPLS algorithm highlight at least as good decoding performance as REW-NPLS algorithm. However, the decoding performance obtained with PREW-NPLS were achieved thanks to sparse models with up to 64% and 75% of the electrodes set to 0 for the left and right hand models respectively using L1-PREW-NPLS. Significance. The designed solution proposed an online incremental adaptive algorithm suitable for online adaptive decoder calibration which estimate sparse decoding solutions. The PREW-NPLS models are suited for portable applications with low computational power using only small number of electrodes with degrading the decoding performance.