In Neural Machine Translation (NMT), each token prediction is conditioned on the source sentence and the target prefix (what has been previously translated at a decoding step). However, previous work on interpretability in NMT has mainly focused solely on source sentence tokens' attributions. Therefore, we lack a full understanding of the influences of every input token (source sentence and target prefix) in the model predictions. In this work, we propose an interpretability method that tracks input tokens' attributions for both contexts. Our method, which can be extended to any encoder-decoder Transformer-based model, allows us to better comprehend the inner workings of current NMT models. We apply the proposed method to both bilingual and multilingual Transformers and present insights into their behaviour.
The state-space models (SSMs) are widely utilized in the analysis of time-series data. SSMs rely on an explicit definition of the state and observation processes. Characterizing these processes is not always easy and becomes a modeling challenge when the dimension of observed data grows or the observed data distribution deviates from the normal distribution. Here, we propose a new formulation of SSM for high-dimensional observation processes. We call this solution the deep direct discriminative decoder (D4). The D4 brings deep neural networks' expressiveness and scalability to the SSM formulation letting us build a novel solution that efficiently estimates the underlying state processes through high-dimensional observation signal. We demonstrate the D4 solutions in simulated and real data such as Lorenz attractors, Langevin dynamics, random walk dynamics, and rat hippocampus spiking neural data and show that the D4 performs better than traditional SSMs and RNNs. The D4 can be applied to a broader class of time-series data where the connection between high-dimensional observation and the underlying latent process is hard to characterize.
With the growing use of machine learning algorithms and ubiquitous sensors, many `perception-to-control' systems are being developed and deployed. To ensure their trustworthiness, improving their robustness through adversarial training is one potential approach. We propose a gradient-free adversarial training technique, named AutoJoin, to effectively and efficiently produce robust models for image-based maneuvering. Compared to other state-of-the-art methods with testing on over 5M images, AutoJoin achieves significant performance increases up to the 40% range against perturbations while improving on clean performance up to 300%. AutoJoin is also highly efficient, saving up to 86% time per training epoch and 90% training data over other state-of-the-art techniques. The core idea of AutoJoin is to use a decoder attachment to the original regression model creating a denoising autoencoder within the architecture. This architecture allows the tasks `maneuvering' and `denoising sensor input' to be jointly learnt and reinforce each other's performance.
In decoding linear block codes, it was shown that noticeable reliability gains can be achieved by introducing learnable parameters to the Belief Propagation (BP) decoder. Despite the success of these methods, there are two key open problems. The first is the lack of interpretation of the learned weights, and the other is the lack of analysis for non-AWGN channels. In this work, we aim to bridge this gap by providing insights into the weights learned and their connection to the structure of the underlying code. We show that the weights are heavily influenced by the distribution of short cycles in the code. We next look at the performance of these decoders in non-AWGN channels, both synthetic and over-the-air channels, and study the complexity vs. performance trade-offs, demonstrating that increasing the number of parameters helps significantly in complex channels. Finally, we show that the decoders with learned weights achieve higher reliability than those with weights optimized analytically under the Gaussian approximation.
In this work, we provide a deterministic alternative to the stochastic variational training of generative autoencoders. We refer to these new generative autoencoders as AutoEncoders within Flows (AEF), since the encoder and decoder are defined as affine layers of an overall invertible architecture. This results in a deterministic encoding of the data, as opposed to the stochastic encoding of VAEs. The paper introduces two related families of AEFs. The first family relies on a partition of the ambient space and is trained by exact maximum-likelihood. The second family exploits a deterministic expansion of the ambient space and is trained by maximizing the log-probability in this extended space. This latter case leaves complete freedom in the choice of encoder, decoder and prior architectures, making it a drop-in replacement for the training of existing VAEs and VAE-style models. We show that these AEFs can have strikingly higher performance than architecturally identical VAEs in terms of log-likelihood and sample quality, especially for low dimensional latent spaces. Importantly, we show that AEF samples are substantially sharper than VAE samples.
Data augmentation methods for Natural Language Processing tasks are explored in recent years, however they are limited and it is hard to capture the diversity on sentence level. Besides, it is not always possible to perform data augmentation on supervised tasks. To address those problems, we propose a neural data augmentation method, which is a combination of Conditional Variational Autoencoder and encoder-decoder Transformer model. While encoding and decoding the input sentence, our model captures the syntactic and semantic representation of the input language with its class condition. Following the developments in the past years on pre-trained language models, we train and evaluate our models on several benchmarks to strengthen the downstream tasks. We compare our method with 3 different augmentation techniques. The presented results show that, our model increases the performance of current models compared to other data augmentation techniques with a small amount of computation power.
Large language models have achieved high performance on various question answering (QA) benchmarks, but the explainability of their output remains elusive. Structured explanations, called entailment trees, were recently suggested as a way to explain and inspect a QA system's answer. In order to better generate such entailment trees, we propose an architecture called Iterative Retrieval-Generation Reasoner (IRGR). Our model is able to explain a given hypothesis by systematically generating a step-by-step explanation from textual premises. The IRGR model iteratively searches for suitable premises, constructing a single entailment step at a time. Contrary to previous approaches, our method combines generation steps and retrieval of premises, allowing the model to leverage intermediate conclusions, and mitigating the input size limit of baseline encoder-decoder models. We conduct experiments using the EntailmentBank dataset, where we outperform existing benchmarks on premise retrieval and entailment tree generation, with around 300% gain in overall correctness.
Modern object detectors have taken the advantages of backbone networks pre-trained on large scale datasets. Except for the backbone networks, however, other components such as the detector head and the feature pyramid network (FPN) remain trained from scratch, which hinders fully tapping the potential of representation models. In this study, we propose to integrally migrate pre-trained transformer encoder-decoders (imTED) to a detector, constructing a feature extraction path which is ``fully pre-trained" so that detectors' generalization capacity is maximized. The essential differences between imTED with the baseline detector are twofold: (1) migrating the pre-trained transformer decoder to the detector head while removing the randomly initialized FPN from the feature extraction path; and (2) defining a multi-scale feature modulator (MFM) to enhance scale adaptability. Such designs not only reduce randomly initialized parameters significantly but also unify detector training with representation learning intendedly. Experiments on the MS COCO object detection dataset show that imTED consistently outperforms its counterparts by $\sim$2.4 AP. Without bells and whistles, imTED improves the state-of-the-art of few-shot object detection by up to 7.6 AP. Code is available at https://github.com/LiewFeng/imTED.
Temporal convolutions have been the paradigm of choice in action segmentation, which enhances long-term receptive fields by increasing convolution layers. However, high layers cause the loss of local information necessary for frame recognition. To solve the above problem, a novel encoder-decoder structure is proposed in this paper, called Cross-Enhancement Transformer. Our approach can be effective learning of temporal structure representation with interactive self-attention mechanism. Concatenated each layer convolutional feature maps in encoder with a set of features in decoder produced via self-attention. Therefore, local and global information are used in a series of frame actions simultaneously. In addition, a new loss function is proposed to enhance the training process that penalizes over-segmentation errors. Experiments show that our framework performs state-of-the-art on three challenging datasets: 50Salads, Georgia Tech Egocentric Activities and the Breakfast dataset.
Convolutional Neural Networks (CNNs) have achieved promising results in medical image segmentation. However, CNNs require lots of training data and are incapable of handling pose and deformation of objects. Furthermore, their pooling layers tend to discard important information such as positions as well as CNNs are sensitive to rotation and affine transformation. Capsule network is a recent new architecture that has achieved better robustness in part-whole representation learning by replacing pooling layers with dynamic routing and convolutional strides, which has shown potential results on popular tasks such as digit classification and object segmentation. In this paper, we propose a 3D encoder-decoder network with Convolutional Capsule Encoder (called 3DConvCaps) to learn lower-level features (short-range attention) with convolutional layers while modeling the higher-level features (long-range dependence) with capsule layers. Our experiments on multiple datasets including iSeg-2017, Hippocampus, and Cardiac demonstrate that our 3D 3DConvCaps network considerably outperforms previous capsule networks and 3D-UNets. We further conduct ablation studies of network efficiency and segmentation performance under various configurations of convolution layers and capsule layers at both contracting and expanding paths.
Ensemble prediction systems are an invaluable tool for weather forecasting. Practically, ensemble predictions are obtained by running several perturbations of the deterministic control forecast. However, ensemble prediction is associated with a high computational cost and often involves statistical post-processing steps to improve its quality. Here we propose to use deep-learning-based algorithms to learn the statistical properties of an ensemble prediction system, the ensemble spread, given only the deterministic control forecast. Thus, once trained, the costly ensemble prediction system will not be needed anymore to obtain future ensemble forecasts, and the statistical properties of the ensemble can be derived from a single deterministic forecast. We adapt the classical pix2pix architecture to a three-dimensional model and also experiment with a shared latent space encoder-decoder model, and train them against several years of operational (ensemble) weather forecasts for the 500 hPa geopotential height. The results demonstrate that the trained models indeed allow obtaining a highly accurate ensemble spread from the control forecast only.
With the COVID-19 global pandemic, computer-assisted diagnoses of medical images have gained much attention, and robust methods of Semantic Segmentation of Computed Tomography (CT) became highly desirable. Semantic Segmentation of CT is one of many research fields of automatic detection of COVID-19 and has been widely explored since the COVID-19 outbreak. In this work, we propose an extensive analysis of how different data augmentation techniques improve the training of encoder-decoder neural networks on this problem. Twenty different data augmentation techniques were evaluated on five different datasets. Each dataset was validated through a five-fold cross-validation strategy, thus resulting in over 3,000 experiments. Our findings show that spatial level transformations are the most promising to improve the learning of neural networks on this problem.
For supervised speech enhancement, contextual information is important for accurate spectral mapping. However, commonly used deep neural networks (DNNs) are limited in capturing temporal contexts. To leverage long-term contexts for tracking a target speaker, this paper treats the speech enhancement as sequence-to-sequence mapping, and propose a novel monaural speech enhancement U-net structure based on Transformer, dubbed U-Former. The key idea is to model long-term correlations and dependencies, which are crucial for accurate noisy speech modeling, through the multi-head attention mechanisms. For this purpose, U-Former incorporates multi-head attention mechanisms at two levels: 1) a multi-head self-attention module which calculate the attention map along both time- and frequency-axis to generate time and frequency sub-attention maps for leveraging global interactions between encoder features, while 2) multi-head cross-attention module which are inserted in the skip connections allows a fine recovery in the decoder by filtering out uncorrelated features. Experimental results illustrate that the U-Former obtains consistently better performance than recent models of PESQ, STOI, and SSNR scores.
Identifying polyps is challenging for automatic analysis of endoscopic images in computer-aided clinical support systems. Models based on convolutional networks (CNN), transformers, and their combinations have been proposed to segment polyps with promising results. However, those approaches have limitations either in modeling the local appearance of the polyps only or lack of multi-level features for spatial dependency in the decoding process. This paper proposes a novel network, namely ColonFormer, to address these limitations. ColonFormer is an encoder-decoder architecture capable of modeling long-range semantic information at both encoder and decoder branches. The encoder is a lightweight architecture based on transformers for modeling global semantic relations at multi scales. The decoder is a hierarchical network structure designed for learning multi-level features to enrich feature representation. Besides, a refinement module is added with a new skip connection technique to refine the boundary of polyp objects in the global map for accurate segmentation. Extensive experiments have been conducted on five popular benchmark datasets for polyp segmentation, including Kvasir, CVC-Clinic DB, CVC-ColonDB, CVC-T, and ETIS-Larib. Experimental results show that our ColonFormer outperforms other state-of-the-art methods on all benchmark datasets.
We are assisting at a growing interest in the development of learning architectures with application to digital communication systems. Herein, we consider the detection/decoding problem. We aim at developing an optimal neural architecture for such a task. The definition of the optimal criterion is a fundamental step. We propose to use the mutual information (MI) of the channel input-output signal pair, which yields to the minimization of the a-posteriori information of the transmitted codeword given the communication channel output observation. The computation of the a-posteriori information is a formidable task, and for the majority of channels it is unknown. Therefore, it has to be learned. For such an objective, we propose a novel neural estimator based on a discriminative formulation. This leads to the derivation of the mutual information neural decoder (MIND). The developed neural architecture is capable not only to solve the decoding problem in unknown channels, but also to return an estimate of the average MI achieved with the coding scheme, as well as the decoding error probability. Several numerical results are reported and compared with maximum a-posteriori and maximum likelihood decoding strategies.
Although generative facial prior and geometric prior have recently demonstrated high-quality results for blind face restoration, producing fine-grained facial details faithful to inputs remains a challenging problem. Motivated by the classical dictionary-based methods and the recent vector quantization (VQ) technique, we propose a VQ-based face restoration method - VQFR. VQFR takes advantage of high-quality low-level feature banks extracted from high-quality faces and can thus help recover realistic facial details. However, the simple application of the VQ codebook cannot achieve good results with faithful details and identity preservation. Therefore, we further introduce two special network designs. 1). We first investigate the compression patch size in the VQ codebook and find that the VQ codebook designed with a proper compression patch size is crucial to balance the quality and fidelity. 2). To further fuse low-level features from inputs while not "contaminating" the realistic details generated from the VQ codebook, we proposed a parallel decoder consisting of a texture decoder and a main decoder. Those two decoders then interact with a texture warping module with deformable convolution. Equipped with the VQ codebook as a facial detail dictionary and the parallel decoder design, the proposed VQFR can largely enhance the restored quality of facial details while keeping the fidelity to previous methods.
We present ViT5, a pretrained Transformer-based encoder-decoder model for the Vietnamese language. With T5-style self-supervised pretraining, ViT5 is trained on a large corpus of high-quality and diverse Vietnamese texts. We benchmark ViT5 on two downstream text generation tasks, Abstractive Text Summarization and Named Entity Recognition. Although Abstractive Text Summarization has been widely studied for the English language thanks to its rich and large source of data, there has been minimal research into the same task in Vietnamese, a much lower resource language. In this work, we perform exhaustive experiments on both Vietnamese Abstractive Summarization and Named Entity Recognition, validating the performance of ViT5 against many other pretrained Transformer-based encoder-decoder models. Our experiments show that ViT5 significantly outperforms existing models and achieves state-of-the-art results on Vietnamese Text Summarization. On the task of Named Entity Recognition, ViT5 is competitive against previous best results from pretrained encoder-based Transformer models. Further analysis shows the importance of context length during the self-supervised pretraining on downstream performance across different settings.
Variational Auto-Encoder (VAE) has become the de-facto learning paradigm in achieving representation learning and generation for natural language at the same time. Nevertheless, existing VAE-based language models either employ elementary RNNs, which is not powerful to handle complex works in the multi-task situation, or fine-tunes two pre-trained language models (PLMs) for any downstream task, which is a huge drain on resources. In this paper, we propose the first VAE framework empowered with adaptive GPT-2s (AdaVAE). Different from existing systems, we unify both the encoder\&decoder of the VAE model using GPT-2s with adaptive parameter-efficient components, and further introduce Latent Attention operation to better construct latent space from transformer models. Experiments from multiple dimensions validate that AdaVAE is competent to effectively organize language in three related tasks (language modeling, representation modeling and guided text generation) even with less than $15\%$ activated parameters in training. Our code is available at \url{https://github.com/ImKeTT/AdaVAE}.
We propose a generative model for text generation, which exhibits disentangled latent representations of syntax and semantics. Contrary to previous work, this model does not need syntactic information such as constituency parses, or semantic information such as paraphrase pairs. Our model relies solely on the inductive bias found in attention-based architectures such as Transformers. In the attention of Transformers, keys handle information selection while values specify what information is conveyed. Our model, dubbed QKVAE, uses Attention in its decoder to read latent variables where one latent variable infers keys while another infers values. We run experiments on latent representations and experiments on syntax/semantics transfer which show that QKVAE displays clear signs of disentangled syntax and semantics. We also show that our model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer. The code for our experiments is publicly available.
In this work, we propose a permutation invariant language model, SymphonyNet, as a solution for symbolic symphony music generation. We propose a novel Multi-track Multi-instrument Repeatable (MMR) representation for symphonic music and model the music sequence using a Transformer-based auto-regressive language model with specific 3-D positional embedding. To overcome length overflow when modeling extra-long symphony tokens, we also propose a modified Byte Pair Encoding algorithm (Music BPE) for music tokens and introduce a novel linear transformer decoder architecture as a backbone. Meanwhile, we train the decoder to learn automatic orchestration as a joint task by masking instrument information from the input. We also introduce a large-scale symbolic symphony dataset for the advance of symphony generation research. Empirical results show that the proposed approach can generate coherent, novel, complex and harmonious symphony as a pioneer solution for multi-track multi-instrument symbolic music generation.