Multilingual neural machine translation (MNMT) trained in multiple language pairs has attracted considerable attention due to fewer model parameters and lower training costs by sharing knowledge among multiple languages. Nonetheless, multilingual training is plagued by language interference degeneration in shared parameters because of the negative interference among different translation directions, especially on high-resource languages. In this paper, we propose the multilingual translation model with the high-resource language-specific training (HLT-MT) to alleviate the negative interference, which adopts the two-stage training with the language-specific selection mechanism. Specifically, we first train the multilingual model only with the high-resource pairs and select the language-specific modules at the top of the decoder to enhance the translation quality of high-resource directions. Next, the model is further trained on all available corpora to transfer knowledge from high-resource languages (HRLs) to low-resource languages (LRLs). Experimental results show that HLT-MT outperforms various strong baselines on WMT-10 and OPUS-100 benchmarks. Furthermore, the analytic experiments validate the effectiveness of our method in mitigating the negative interference in multilingual training.
Drawing from memory the face of a friend you have not seen in years is a difficult task. However, if you happen to cross paths, you would easily recognize each other. The biological memory is equipped with an impressive compression algorithm that can store the essential, and then infer the details to match perception. The Willshaw Memory is a simple abstract model for cortical computations which implements mechanisms of biological memories. Using our recently proposed sparse coding prescription for visual patterns, this model can store and retrieve an impressive amount of real-world data in a fault-tolerant manner. In this paper, we extend the capabilities of the basic Associative Memory Model by using a Multiple-Modality framework. In this setting, the memory stores several modalities (e.g., visual, or textual) of each pattern simultaneously. After training, the memory can be used to infer missing modalities when just a subset is perceived. Using a simple encoder-memory-decoder architecture, and a newly proposed iterative retrieval algorithm for the Willshaw Model, we perform experiments on the MNIST dataset. By storing both the images and labels as modalities, a single Memory can be used not only to retrieve and complete patterns but also to classify and generate new ones. We further discuss how this model could be used for other learning tasks, thus serving as a biologically-inspired framework for learning.
Radiology report generation (RRG) aims to describe automatically a radiology image with human-like language and could potentially support the work of radiologists, reducing the burden of manual reporting. Previous approaches often adopt an encoder-decoder architecture and focus on single-modal feature learning, while few studies explore cross-modal feature interaction. Here we propose a Cross-modal PROtotype driven NETwork (XPRONET) to promote cross-modal pattern learning and exploit it to improve the task of radiology report generation. This is achieved by three well-designed, fully differentiable and complementary modules: a shared cross-modal prototype matrix to record the cross-modal prototypes; a cross-modal prototype network to learn the cross-modal prototypes and embed the cross-modal information into the visual and textual features; and an improved multi-label contrastive loss to enable and enhance multi-label prototype learning. XPRONET obtains substantial improvements on the IU-Xray and MIMIC-CXR benchmarks, where its performance exceeds recent state-of-the-art approaches by a large margin on IU-Xray and comparable performance on MIMIC-CXR.
Cursive handwritten text recognition is a challenging research problem in the domain of pattern recognition. The current state-of-the-art approaches include models based on convolutional recurrent neural networks and multi-dimensional long short-term memory recurrent neural networks techniques. These methods are highly computationally extensive as well model is complex at design level. In recent studies, combination of convolutional neural network and gated convolutional neural networks based models demonstrated less number of parameters in comparison to convolutional recurrent neural networks based models. In the direction to reduced the total number of parameters to be trained, in this work, we have used depthwise convolution in place of standard convolutions with a combination of gated-convolutional neural network and bidirectional gated recurrent unit to reduce the total number of parameters to be trained. Additionally, we have also included a lexicon based word beam search decoder at testing step. It also helps in improving the the overall accuracy of the model. We have obtained 3.84% character error rate and 9.40% word error rate on IAM dataset; 4.88% character error rate and 14.56% word error rate in George Washington dataset, respectively.
Attention-based models such as transformers have shown outstanding performance on dense prediction tasks, such as semantic segmentation, owing to their capability of capturing long-range dependency in an image. However, the benefit of transformers for monocular depth prediction has seldom been explored so far. This paper benchmarks various transformer-based models for the depth estimation task on an indoor NYUV2 dataset and an outdoor KITTI dataset. We propose a novel attention-based architecture, Depthformer for monocular depth estimation that uses multi-head self-attention to produce the multiscale feature maps, which are effectively combined by our proposed decoder network. We also propose a Transbins module that divides the depth range into bins whose center value is estimated adaptively per image. The final depth estimated is a linear combination of bin centers for each pixel. Transbins module takes advantage of the global receptive field using the transformer module in the encoding stage. Experimental results on NYUV2 and KITTI depth estimation benchmark demonstrate that our proposed method improves the state-of-the-art by 3.3%, and 3.3% respectively in terms of Root Mean Squared Error (RMSE). Code is available at https://github.com/ashutosh1807/Depthformer.git.
The Synthetic Control method has pioneered a class of powerful data-driven techniques to estimate the counterfactual reality of a unit from donor units. At its core, the technique involves a linear model fitted on the pre-intervention period that combines donor outcomes to yield the counterfactual. However, linearly combining spatial information at each time instance using time-agnostic weights fails to capture important inter-unit and intra-unit temporal contexts and complex nonlinear dynamics of real data. We instead propose an approach to use local spatiotemporal information before the onset of the intervention as a promising way to estimate the counterfactual sequence. To this end, we suggest a Transformer model that leverages particular positional embeddings, a modified decoder attention mask, and a novel pre-training task to perform spatiotemporal sequence-to-sequence modeling. Our experiments on synthetic data demonstrate the efficacy of our method in the typical small donor pool setting and its robustness against noise. We also generate actionable healthcare insights at the population and patient levels by simulating a state-wide public health policy to evaluate its effectiveness, an in silico trial for asthma medications to support randomized controlled trials, and a medical intervention for patients with Friedreich's ataxia to improve clinical decision-making and promote personalized therapy.
We explore a new idea for learning based shape reconstruction from a point cloud, based on the recently popularized implicit neural shape representations. We cast the problem as a few-shot learning of implicit neural signed distance functions in feature space, that we approach using gradient based meta-learning. We use a convolutional encoder to build a feature space given the input point cloud. An implicit decoder learns to predict signed distance values given points represented in this feature space. Setting the input point cloud, i.e. samples from the target shape function's zero level set, as the support (i.e. context) in few-shot learning terms, we train the decoder such that it can adapt its weights to the underlying shape of this context with a few (5) tuning steps. We thus combine two types of implicit neural network conditioning mechanisms simultaneously for the first time, namely feature encoding and meta-learning. Our numerical and qualitative evaluation shows that in the context of implicit reconstruction from a sparse point cloud, our proposed strategy, i.e. meta-learning in feature space, outperforms existing alternatives, namely standard supervised learning in feature space, and meta-learning in euclidean space, while still providing fast inference.
Deep learning (DL) in general and Recurrent neural networks (RNNs) in particular have seen high success levels in sequence based applications. This paper pertains to RNNs for time series modelling and forecasting. We propose a novel RNN architecture capturing (stochastic) seasonal correlations intelligently while capable of accurate multi-step forecasting. It is motivated from the well-known encoder-decoder (ED) architecture and multiplicative seasonal auto-regressive model. It incorporates multi-step (multi-target) learning even in the presence (or absence) of exogenous inputs. It can be employed on single or multiple sequence data. For the multiple sequence case, we also propose a novel greedy recursive procedure to build (one or more) predictive models across sequences when per-sequence data is less. We demonstrate via extensive experiments the utility of our proposed architecture both in single sequence and multiple sequence scenarios.
In this work we introduce a biologically inspired long-range skip connection for the UNet architecture that relies on the perceptual illusion of hybrid images, being images that simultaneously encode two images. The fusion of early encoder features with deeper decoder ones allows UNet models to produce finer-grained dense predictions. While proven in segmentation tasks, the network's benefits are down-weighted for dense regression tasks as these long-range skip connections additionally result in texture transfer artifacts. Specifically for depth estimation, this hurts smoothness and introduces false positive edges which are detrimental to the task due to the depth maps' piece-wise smooth nature. The proposed HybridSkip connections show improved performance in balancing the trade-off between edge preservation, and the minimization of texture transfer artifacts that hurt smoothness. This is achieved by the proper and balanced exchange of information that Hybrid-Skip connections offer between the high and low frequency, encoder and decoder features, respectively.
Intermediate layer output (ILO) regularization by means of multitask training on encoder side has been shown to be an effective approach to yielding improved results on a wide range of end-to-end ASR frameworks. In this paper, we propose a novel method to do ILO regularized training differently. Instead of using conventional multitask methods that entail more training overhead, we directly make the intermediate layer output as input to the decoder, that is, our decoder not only accepts the output of the final encoder layer as input, it also takes the output of the encoder ILO as input during training. With the proposed method, as both encoder and decoder are simultaneously "regularized", the network is more sufficiently trained, consistently leading to improved results, over the ILO-based CTC method, as well as over the original attention-based modeling method without the proposed method employed.
The Transformer-based encoder-decoder architecture has recently made significant advances in recognizing handwritten mathematical expressions. However, the transformer model still suffers from the lack of coverage problem, making its expression recognition rate (ExpRate) inferior to its RNN counterpart. Coverage information, which records the alignment information of the past steps, has proven effective in the RNN models. In this paper, we propose CoMER, a model that adopts the coverage information in the transformer decoder. Specifically, we propose a novel Attention Refinement Module (ARM) to refine the attention weights with past alignment information without hurting its parallelism. Furthermore, we take coverage information to the extreme by proposing self-coverage and cross-coverage, which utilize the past alignment information from the current and previous layers. Experiments show that CoMER improves the ExpRate by 0.61%/2.09%/1.59% compared to the current state-of-the-art model, and reaches 59.33%/59.81%/62.97% on the CROHME 2014/2016/2019 test sets.
The purpose of this contribution is to introduce a new method of signal prediction in video coding. Unlike most existent prediction methods that either use temporal or use spatial correlations to generate the prediction signal, the proposed method uses spatial and temporal correlations at the same time. The spatio-temporal prediction is obtained by first performing motion compensation for a macroblock, followed by a refinement step that pays attention to the correlations between the macroblock and its surroundings. At the decoder, the refinement step can be performed in the same manner, thus no additional side information has to be transmitted. Implementation of the spatial refinement step into the H.264/AVC video codec leads to reduction in data rate of up to nearly 15% and increase in PSNR of up to 0.75 dB, compared to pure motion compensated prediction.
In this paper, we propose a novel framework for music-driven dance motion synthesis with controllable key pose constraint. In contrast to methods that generate dance motion sequences only based on music without any other controllable conditions, this work targets on synthesizing high-quality dance motion driven by music as well as customized poses performed by users. Our model involves two single-modal transformer encoders for music and motion representations and a cross-modal transformer decoder for dance motions generation. The cross-modal transformer decoder achieves the capability of synthesizing smooth dance motion sequences, which keeps a consistency with key poses at corresponding positions, by introducing the local neighbor position embedding. Such mechanism makes the decoder more sensitive to key poses and the corresponding positions. Our dance synthesis model achieves satisfactory performance both on quantitative and qualitative evaluations with extensive experiments, which demonstrates the effectiveness of our proposed method.
Large, pretrained language models infer powerful representations that encode rich semantic and syntactic content, albeit implicitly. In this work we introduce a novel neural language model that enforces, via inductive biases, explicit relational structures which allow for compositionality onto the output representations of pretrained language models. Specifically, the model encodes sentences into sequences of symbols (composed representations), which correspond to the nodes visited by biased random walkers on a global latent graph, and infers the posterior distribution of the latter. We first demonstrate that the model is able to uncover ground-truth graphs from artificially generated datasets of random token sequences. Next, we leverage pretrained BERT and GPT-2 language models as encoder and decoder, respectively, to infer networks of symbols (schemata) from natural language datasets. Our experiments show that (i) the inferred symbols can be interpreted as encoding different aspects of language, as e.g. topics or sentiments, and that (ii) GPT-like models can effectively be conditioned on symbolic representations. Finally, we explore training autoregressive, random walk ``reasoning" models on schema networks inferred from commonsense knowledge databases, and using the sampled paths to enhance the performance of pretrained language models on commonsense If-Then reasoning tasks.
In this work, we present an end-to-end binaural speech synthesis system that combines a low-bitrate audio codec with a powerful binaural decoder that is capable of accurate speech binauralization while faithfully reconstructing environmental factors like ambient noise or reverb. The network is a modified vector-quantized variational autoencoder, trained with several carefully designed objectives, including an adversarial loss. We evaluate the proposed system on an internal binaural dataset with objective metrics and a perceptual study. Results show that the proposed approach matches the ground truth data more closely than previous methods. In particular, we demonstrate the capability of the adversarial loss in capturing environment effects needed to create an authentic auditory scene.
A convolutional autoencoder is trained using a database of airfoil aerodynamic simulations and assessed in terms of overall accuracy and interpretability. The goal is to predict the stall and to investigate the ability of the autoencoder to distinguish between the linear and non-linear response of the airfoil pressure distribution to changes in the angle of attack. After a sensitivity analysis on the learning infrastructure, we investigate the latent space identified by the autoencoder targeting extreme compression rates, i.e. very low-dimensional reconstructions. We also propose a strategy to use the decoder to generate new synthetic airfoil geometries and aerodynamic solutions by interpolation and extrapolation in the latent representation learned by the autoencoder.
Neural networks have proven to be a formidable tool to tackle the problem of speech coding at very low bit rates. However, the design of a neural coder that can be operated robustly under real-world conditions remains a major challenge. Therefore, we present Neural End-2-End Speech Codec (NESC) a robust, scalable end-to-end neural speech codec for high-quality wideband speech coding at 3 kbps. The encoder uses a new architecture configuration, which relies on our proposed Dual-PathConvRNN (DPCRNN) layer, while the decoder architecture is based on our previous work Streamwise-StyleMelGAN. Our subjective listening tests on clean and noisy speech show that NESC is particularly robust to unseen conditions and signal perturbations.
This paper studies multi-task training of retrieval-augmented generation models for knowledge-intensive tasks. We propose to clean the training set by utilizing a distinct property of knowledge-intensive generation: The connection of query-answer pairs to items in the knowledge base. We filter training examples via a threshold of confidence on the relevance labels, whether a pair is answerable by the knowledge base or not. We train a single Fusion-in-Decoder (FiD) generator on seven combined tasks of the KILT benchmark. The experimental results suggest that our simple yet effective approach substantially improves competitive baselines on two strongly imbalanced tasks; and shows either smaller improvements or no significant regression on the remaining tasks. Furthermore, we demonstrate our multi-task training with relevance label sampling scales well with increased model capacity and achieves state-of-the-art results in five out of seven KILT tasks.
Self-supervised learning methods overcome the key bottleneck for building more capable AI: limited availability of labeled data. However, one of the drawbacks of self-supervised architectures is that the representations that they learn are implicit and it is hard to extract meaningful information about the encoded world states, such as 3D structure of the visual scene encoded in a depth map. Moreover, in the visual domain such representations only rarely undergo evaluations that may be critical for downstream tasks, such as vision for autonomous cars. Herein, we propose a framework for evaluating visual representations for illumination invariance in the context of depth perception. We develop a new predictive coding-based architecture and a hybrid fully-supervised/self-supervised learning method. We propose a novel architecture that extends the predictive coding approach: PRedictive Lateral bottom-Up and top-Down Encoder-decoder Network (PreludeNet), which explicitly learns to infer and predict depth from video frames. In PreludeNet, the encoder's stack of predictive coding layers is trained in a self-supervised manner, while the predictive decoder is trained in a supervised manner to infer or predict the depth. We evaluate the robustness of our model on a new synthetic dataset, in which lighting conditions (such as overall illumination, and effect of shadows) can be be parametrically adjusted while keeping all other aspects of the world constant. PreludeNet achieves both competitive depth inference performance and next frame prediction accuracy. We also show how this new network architecture, coupled with the hybrid fully-supervised/self-supervised learning method, achieves balance between the said performance and invariance to changes in lighting. The proposed framework for evaluating visual representations can be extended to diverse task domains and invariance tests.
Videos are prone to tampering attacks that alter the meaning and deceive the audience. Previous video forgery detection schemes find tiny clues to locate the tampered areas. However, attackers can successfully evade supervision by destroying such clues using video compression or blurring. This paper proposes a video watermarking network for tampering localization. We jointly train a 3D-UNet-based watermark embedding network and a decoder that predicts the tampering mask. The perturbation made by watermark embedding is close to imperceptible. Considering that there is no off-the-shelf differentiable video codec simulator, we propose to mimic video compression by ensembling simulation results of other typical attacks, e.g., JPEG compression and blurring, as an approximation. Experimental results demonstrate that our method generates watermarked videos with good imperceptibility and robustly and accurately locates tampered areas within the attacked version.