Learning joint embedding space for various modalities is of vital importance for multimodal fusion. Mainstream modality fusion approaches fail to achieve this goal, leaving a modality gap which heavily affects cross-modal fusion. In this paper, we propose a novel adversarial encoder-decoder-classifier framework to learn a modality-invariant embedding space. Since the distributions of various modalities vary in nature, to reduce the modality gap, we translate the distributions of source modalities into that of target modality via their respective encoders using adversarial training. Furthermore, we exert additional constraints on embedding space by introducing reconstruction loss and classification loss. Then we fuse the encoded representations using hierarchical graph neural network which explicitly explores unimodal, bimodal and trimodal interactions in multi-stage. Our method achieves state-of-the-art performance on multiple datasets. Visualization of the learned embeddings suggests that the joint embedding space learned by our method is discriminative. code is available at: \url{https://github.com/TmacMai/ARGF_multimodal_fusion}
Consistency is one of the major challenges faced by dialogue agents. A human-like dialogue agent should not only respond naturally, but also maintain a consistent persona. In this paper, we exploit the advantages of natural language inference (NLI) technique to address the issue of generating persona consistent dialogues. Different from existing work that re-ranks the retrieved responses through an NLI model, we cast the task as a reinforcement learning problem and propose to exploit the NLI signals from response-persona pairs as rewards for the process of dialogue generation. Specifically, our generator employs an attention-based encoder-decoder to generate persona-based responses. Our evaluator consists of two components: an adversarially trained naturalness module and an NLI based consistency module. Moreover, we use another well-performed NLI model in the evaluation of persona-consistency. Experimental results on both human and automatic metrics, including the model-based consistency evaluation, demonstrate that the proposed approach outperforms strong generative baselines, especially in the persona-consistency of generated responses.
Many successful learning targets such as minimizing dice loss and cross-entropy loss have enabled unprecedented breakthroughs in segmentation tasks. Beyond these semantic metrics, this paper aims to introduce location supervision into semantic segmentation. Based on this idea, we present a Location-aware Upsampling (LaU) that adaptively refines the interpolating coordinates with trainable offsets. Then, location-aware losses are established by encouraging pixels to move towards well-classified locations. An LaU is offset prediction coupled with interpolation, which is trained end-to-end to generate confidence score at each position from coarse to fine. Guided by location-aware losses, the new module can replace its plain counterpart (\textit{e.g.}, bilinear upsampling) in a plug-and-play manner to further boost the leading encoder-decoder approaches. Extensive experiments validate the consistent improvement over the state-of-the-art methods on benchmark datasets. Our code is available at https://github.com/HolmesShuan/Location-aware-Upsampling-for-Semantic-Segmentation
This paper studies the prediction of chord progressions for jazz music by relying on machine learning models. The motivation of our study comes from the recent success of neural networks for performing automatic music composition. Although high accuracies are obtained in single-step prediction scenarios, most models fail to generate accurate multi-step chord predictions. In this paper, we postulate that this comes from the multi-scale structure of musical information and propose new architectures based on an iterative temporal aggregation of input labels. Specifically, the input and ground truth labels are merged into increasingly large temporal bags, on which we train a family of encoder-decoder networks for each temporal scale. In a second step, we use these pre-trained encoder bottleneck features at each scale in order to train a final encoder-decoder network. Furthermore, we rely on different reductions of the initial chord alphabet into three adapted chord alphabets. We perform evaluations against several state-of-the-art models and show that our multi-scale architecture outperforms existing methods in terms of accuracy and perplexity, while requiring relatively few parameters. We analyze musical properties of the results, showing the influence of downbeat position within the analysis window on accuracy, and evaluate errors using a musically-informed distance metric.
Understanding and monitoring the complex and dynamic processes of the Sun is important for a number of human activities on Earth and in space. For this reason, NASA's Solar Dynamics Observatory (SDO) has been continuously monitoring the multi-layered Sun's atmosphere in high-resolution since its launch in 2010, generating terabytes of observational data every day. The synergy between machine learning and this enormous amount of data has the potential, still largely unexploited, to advance our understanding of the Sun and extend the capabilities of heliophysics missions. In the present work, we show that deep learning applied to SDO data can be successfully used to create a high-fidelity virtual telescope that generates synthetic observations of the solar corona by image translation. Towards this end we developed a deep neural network, structured as an encoder-decoder with skip connections (U-Net), that reconstructs the Sun's image of one instrument channel given temporally aligned images in three other channels. The approach we present has the potential to reduce the telemetry needs of SDO, enhance the capabilities of missions that have less observing channels, and transform the concept development of future missions.
Recent works in dialogue state tracking (DST) focus on an open vocabulary-based setting to resolve scalability and generalization issues of the predefined ontology-based approaches. However, they are inefficient in that they predict the dialogue state at every turn from scratch. Here, we consider dialogue state as an explicit fixed-sized memory and propose a selectively overwriting mechanism for more efficient DST. This mechanism consists of two steps: (1) predicting state operation on each of the memory slots, and (2) overwriting the memory with new values, of which only a few are generated according to the predicted state operations. Our method decomposes DST into two sub-tasks and guides the decoder to focus only on one of the tasks, thus reducing the burden of the decoder. This enhances the effectiveness of training and DST performance. Our SOM-DST (Selectively Overwriting Memory for Dialogue State Tracking) model achieves state-of-the-art joint goal accuracy with 51.72% in MultiWOZ 2.0 and 53.01% in MultiWOZ 2.1 in an open vocabulary-based DST setting. In addition, we analyze the accuracy gaps between the current and the ground truth-given situations and suggest that it is a promising direction to improve state operation prediction to boost the DST performance.
Designing codes that combat the noise in a communication medium has remained a significant area of research in information theory as well as wireless communications. Asymptotically optimal channel codes have been developed by mathematicians for communicating under canonical models after over 60 years of research. On the other hand, in many non-canonical channel settings, optimal codes do not exist and the codes designed for canonical models are adapted via heuristics to these channels and are thus not guaranteed to be optimal. In this work, we make significant progress on this problem by designing a fully end-to-end jointly trained neural encoder and decoder, namely, Turbo Autoencoder (TurboAE), with the following contributions: ($a$) under moderate block lengths, TurboAE approaches state-of-the-art performance under canonical channels; ($b$) moreover, TurboAE outperforms the state-of-the-art codes under non-canonical settings in terms of reliability. TurboAE shows that the development of channel coding design can be automated via deep learning, with near-optimal performance.
Despite the success of style transfer in image processing, it has seen limited progress in natural language generation. Part of the problem is that content is not as easily decoupled from style in the text domain. Curiously, in the field of stylometry, content does not figure prominently in practical methods of discriminating stylistic elements, such as authorship and genre. Rather, syntax and function words are the most salient features. Drawing on this work, we model style as a suite of low-level linguistic controls, such as frequency of pronouns, prepositions, and subordinate clause constructions. We train a neural encoder-decoder model to reconstruct reference sentences given only content words and the setting of the controls. We perform style transfer by keeping the content words fixed while adjusting the controls to be indicative of another style. In experiments, we show that the model reliably responds to the linguistic controls and perform both automatic and manual evaluations on style transfer. We find we can fool a style classifier 84% of the time, and that our model produces highly diverse and stylistically distinctive outputs. This work introduces a formal, extendable model of style that can add control to any neural text generation system.
Deep neural networks (DNN) are quickly becoming the de facto standard modeling method for many natural language generation (NLG) tasks. In order for such models to truly be useful, they must be capable of correctly generating utterances for novel meaning representations (MRs) at test time. In practice, even sophisticated DNNs with various forms of semantic control frequently fail to generate utterances faithful to the input MR. In this paper, we propose an architecture agnostic self-training method to sample novel MR/text utterance pairs to augment the original training data. Remarkably, after training on the augmented data, even simple encoder-decoder models with greedy decoding are capable of generating semantically correct utterances that are as good as state-of-the-art outputs in both automatic and human evaluations of quality.
The primary goal of skeletal motion prediction is to generate future motion by observing a sequence of 3D skeletons. A key challenge in motion prediction is the fact that a motion can often be performed in several different ways, with each consisting of its own configuration of poses and their spatio-temporal dependencies, and as a result, the predicted poses often converge to the motionless poses or non-human like motions in long-term prediction. This leads us to define a hierarchical recurrent network model that explicitly characterizes these internal configurations of poses and their local and global spatio-temporal dependencies. The model introduces a latent vector variable from the Lie algebra to represent spatial and temporal relations simultaneously. Furthermore, a structured stack LSTM-based decoder is devised to decode the predicted poses with a new loss function defined to estimate the quantized weight of each body part in a pose. Empirical evaluations on benchmark datasets suggest our approach significantly outperforms the state-of-the-art methods on both short-term and long-term motion prediction.
In this paper, we present a multi-class eye segmentation method that can run the hardware limitations for real-time inference. Our approach includes three major stages: get a grayscale image from the input, segment three distinct eye region with a deep network, and remove incorrect areas with heuristic filters. Our model based on the encoder decoder structure with the key is the depthwise convolution operation to reduce the computation cost. We experiment on OpenEDS, a large scale dataset of eye images captured by a head-mounted display with two synchronized eye facing cameras. We achieved the mean intersection over union (mIoU) of 94.85% with a model of size 0.4 megabytes. The source code are available https://github.com/th2l/Eye_VR_Segmentation
Taking greedy decoding algorithm as it should be, this work focuses on further strengthening the model itself for Chinese word segmentation (CWS), which results in an even more fast and more accurate CWS model. Our model consists of an attention only stacked encoder and a light enough decoder for the greedy segmentation plus two highway connections for smoother training, in which the encoder is composed of a newly proposed Transformer variant, Gaussian-masked Directional (GD) Transformer, and a biaffine attention scorer. With the effective encoder design, our model only needs to take unigram features for scoring. Our model is evaluated on SIGHAN Bakeoff benchmark datasets. The experimental results show that with the highest segmentation speed, the proposed model achieves new state-of-the-art or comparable performance against strong baselines in terms of strict closed test setting.
The emergence of various intelligent mobile applications demands the deployment of powerful deep learning models at resource-constrained mobile devices. The device-edge co-inference framework provides a promising solution by splitting a neural network at a mobile device and an edge computing server. In order to balance the on-device computation and the communication overhead, the splitting point needs to be carefully picked, while the intermediate feature needs to be compressed before transmission. Existing studies decoupled the design of model splitting, feature compression, and communication, which may lead to excessive resource consumption of the mobile device. In this paper, we introduce an end-to-end architecture, named BottleNet++, that consists of an encoder, a non-trainable channel layer, and a decoder for more efficient feature compression and transmission. The encoder and decoder essentially implement joint source-channel coding via convolutional neural networks (CNNs), while explicitly considering the effect of channel noise. By exploiting the strong sparsity and the fault-tolerant property of the intermediate feature in a deep neural network (DNN), BottleNet++ achieves a much higher compression ratio than existing methods. Furthermore, by providing the channel condition to the encoder as an input, our method enjoys a strong generalization ability in different channel conditions. Compared with merely transmitting intermediate data without feature compression, BottleNet++ achieves up to 64x bandwidth reduction over the additive white Gaussian noise channel and up to 256x bit compression ratio in the binary erasure channel, with less than 2% reduction in accuracy. With a higher compression ratio, BottleNet++ enables splitting a DNN at earlier layers, which leads to up to 3x reduction in on-device computation compared with other compression methods.
Interest in the research areas related to meme propagation and generation has been increasing rapidly in the last couple of years. Meme datasets available online are either specific to a context or contain no class information. Here, we prepare a large-scale dataset of memes with captions and class labels. The dataset consists of 1.1 million meme captions from 128 classes. We also provide reasoning for the existence of broad categories, called "themes" across the meme dataset; each theme consists of multiple meme classes. Our generation system uses a trained state-of-the-art transformer-based model for caption generation by employing an encoder-decoder architecture. We develop a web interface, called Memeify for users to generate memes of their choice, and explain in detail, the working of individual components of the system. We also perform a qualitative evaluation of the generated memes by conducting a user study. A link to the demonstration of the Memeify system is https://youtu.be/P_Tfs0X-czs.
In this paper, we investigate the benefit that off-the-shelf word embedding can bring to the sequence-to-sequence (seq-to-seq) automatic speech recognition (ASR). We first introduced the word embedding regularization by maximizing the cosine similarity between a transformed decoder feature and the target word embedding. Based on the regularized decoder, we further proposed the fused decoding mechanism. This allows the decoder to consider the semantic consistency during decoding by absorbing the information carried by the transformed decoder feature, which is learned to be close to the target word embedding. Initial results on LibriSpeech demonstrated that pre-trained word embedding can significantly lower ASR recognition error with a negligible cost, and the choice of word embedding algorithms among Skip-gram, CBOW and BERT is important.
Attribution editing has achieved remarkable progress in recent years owing to the encoder-decoder structure and generative adversarial network (GAN). However, it remains challenging in generating high-quality images with accurate attribute transformation. Attacking these problems, the work proposes a novel selective attribute editing model based on classification adversarial network (referred to as ClsGAN) that shows good balance between attribute transfer accuracy and photo-realistic images. Considering that the editing images are prone to be affected by original attribute due to skip-connection in encoder-decoder structure, an upper convolution residual network (referred to as Tr-resnet) is presented to selectively extract information from the source image and target label. In addition, to further improve the transfer accuracy of generated images, an attribute adversarial classifier (referred to as Atta-cls) is introduced to guide the generator from the perspective of attribute through learning the defects of attribute transfer images. Experimental results on CelebA demonstrate that our ClsGAN performs favorably against state-of-the-art approaches in image quality and transfer accuracy. Moreover, ablation studies are also designed to verify the great performance of Tr-resnet and Atta-cls.
Classical supervised classification tasks search for a nonlinear mapping that maps each encoded feature directly to a probability mass over the labels. Such a learning framework typically lacks the intuition that encoded features from the same class tend to be similar and thus has little interpretability for the learned features. In this paper, we propose a novel supervised learning model named Supervised-Encoding Quantizer (SEQ). The SEQ applies a quantizer to cluster and classify the encoded features. We found that the quantizer provides an interpretable graph where each cluster in the graph represents a class of data samples that have a particular style. We also trained a decoder that can decode convex combinations of the encoded features from similar and different clusters and provide guidance on style transfer between sub-classes.
This extended abstract describes our solution for the Traffic4Cast Challenge 2019. The key problem we addressed is to properly model both low-level (pixel based) and high-level spatial information while still preserve the temporal relations among the frames. Our approach is inspired by the recent adoption of convolutional features into a recurrent neural networks such as LSTM to jointly capture the spatio-temporal dependency. While this approach has been proven to surpass the traditional stacked CNNs (using 2D or 3D kernels) in action recognition, we observe suboptimal performance in traffic prediction setting. Therefore, we apply a number of adaptations in the frame encoder-decoder layers and in sampling procedure to better capture the high-resolution trajectories, and to increase the training efficiency.
The real-time segmentation of surgical instruments plays a crucial role in robot-assisted surgery. However, it is still a challenging task to implement deep learning models to do real-time segmentation for surgical instruments due to their high computational costs and slow inference speed. In this paper, we propose an attention-guided lightweight network (LWANet), which can segment surgical instruments in real-time. LWANet adopts encoder-decoder architecture, where the encoder is the lightweight network MobileNetV2, and the decoder consists of depthwise separable convolution, attention fusion block, and transposed convolution. Depthwise separable convolution is used as the basic unit to construct the decoder, which can reduce the model size and computational costs. Attention fusion block captures global contexts and encodes semantic dependencies between channels to emphasize target regions, contributing to locating the surgical instrument. Transposed convolution is performed to upsample feature maps for acquiring refined edges. LWANet can segment surgical instruments in real-time while takes little computational costs. Based on 960*544 inputs, its inference speed can reach 39 fps with only 3.39 GFLOPs. Also, it has a small model size and the number of parameters is only 2.06 M. The proposed network is evaluated on two datasets. It achieves state-of-the-art performance 94.10% mean IOU on Cata7 and obtains a new record on EndoVis 2017 with a 4.10% increase on mean IOU.
While deep learning has received a surge of interest in a variety of fields in recent years, major deep learning models barely use complex numbers. However, speech, signal and audio data are naturally complex-valued after Fourier Transform, and studies have shown a potentially richer representation of complex nets. In this paper, we propose a Complex Transformer, which incorporates the transformer model as a backbone for sequence modeling; we also develop attention and encoder-decoder network operating for complex input. The model achieves state-of-the-art performance on the MusicNet dataset and an In-phase Quadrature (IQ) signal dataset.