Abstract--We propose a joint source and channel coding (JSCC) technique for wireless image transmission that does not rely on explicit codes for either compression or error correction; instead, it directly maps the image pixel values to the real/complex - valued channel input symbols. We parameterize the encoder and decoder functions by two convolutional neural networks (CNNs), which are trained jointly, and can be considered as an autoencoder with a nontrainable layer in the middle that represents the noisy communication channel. Our results show that the proposed deep JSCC scheme outperforms digital transmission concatenating JPEG or JPEG2000 compression with a capacity achieving channel code at low signal-tonoise ratio (SNR) and channel bandwidth values in the presence of additive white Gaussian noise. More strikingly, deep JSCC does not suffer from the "cliff effect", and it provides a graceful performance degradation as the channel SNR varies with respect to the training SNR. In the case of a slow Rayleigh fading channel, deep JSCC can learn to communicate without explicit pilot signals or channel estimation, and significantly outperforms separation-based digital communication at all SNR and channel bandwidth values. I. Introduction Modern communication systems employ a two step encoding process for the transmission of image/video data (see Figure 1a for an illustration): (i) the image/video data is first compressed with a source coding algorithm in order to get rid of the inherent redundancy, and to reduce the amount of transferred information; and (ii) the compressed bitstream is first encoded with an error correcting code, which enables resilient transmission against errors, and then modulated. Shannon's separation theorem proves that this two-step source and channel coding approach is optimal theoretically in the asymptotic limit of infinitely E. Bourtsoulatze is with the Communications and Information Systems Group, Department of Electronic and Electrical Engineering, University College London, London, UK.

Recent advances in deep learning have brought to the fore models that can make multiple computational steps in the service of completing a task; these are capable of describ- ing long-term dependencies in sequential data. Novel recurrent attention models over possibly large external memory modules constitute the core mechanisms that enable these capabilities. Our work addresses learning subtler and more complex underlying temporal dynamics in language modeling tasks that deal with sparse sequential data. To this end, we improve upon these recent advances, by adopting concepts from the field of Bayesian statistics, namely variational inference. Our proposed approach consists in treating the network parameters as latent variables with a prior distribution imposed over them. Our statistical assumptions go beyond the standard practice of postulating Gaussian priors. Indeed, to allow for handling outliers, which are prevalent in long observed sequences of multivariate data, multivariate t-exponential distributions are imposed. On this basis, we proceed to infer corresponding posteriors; these can be used for inference and prediction at test time, in a way that accounts for the uncertainty in the available sparse training data. Specifically, to allow for our approach to best exploit the merits of the t-exponential family, our method considers a new t-divergence measure, which generalizes the concept of the Kullback-Leibler divergence. We perform an extensive experimental evaluation of our approach, using challenging language modeling benchmarks, and illustrate its superiority over existing state-of-the-art techniques.

While current state-of-the-art NMT models, such as RNN seq2seq and Transformers, possess a large number of parameters, they are still shallow in comparison to convolutional models used for both text and vision applications. In this work we attempt to train significantly (2-3x) deeper Transformer and Bi-RNN encoders for machine translation. We propose a simple modification to the attention mechanism that eases the optimization of deeper models, and results in consistent gains of 0.7-1.1 BLEU on the benchmark WMT'14 English-German and WMT'15 Czech-English tasks for both architectures.

We introduce Texar, an open-source toolkit aiming to support the broad set of text generation tasks that transforms any inputs into natural language, such as machine translation, summarization, dialog, content manipulation, and so forth. With the design goals of modularity, versatility, and extensibility in mind, Texar extracts common patterns underlying the diverse tasks and methodologies, creates a library of highly reusable modules and functionalities, and allows arbitrary model architectures and algorithmic paradigms. In Texar, model architecture, losses, and learning processes are fully decomposed. Modules at high concept level can be freely assembled or plugged in/swapped out. These features make Texar particularly suitable for researchers and practitioners to do fast prototyping and experimentation, as well as foster technique sharing across different text generation tasks. We provide case studies to demonstrate the use and advantage of the toolkit. Texar is released under Apache license 2.0 at https://github.com/asyml/texar.

We propose KDSL, a new word sense disambiguation (WSD) framework that utilizes knowledge to automatically generate sense-labeled data for supervised learning. First, from WordNet, we automatically construct a semantic knowledge base called DisDict, which provides refined feature words that highlight the differences among word senses, i.e., synsets. Second, we automatically generate new sense-labeled data by DisDict from unlabeled corpora. Third, these generated data, together with manually labeled data and unlabeled data, are fed to a neural framework conducting supervised and unsupervised learning jointly to model the semantic relations among synsets, feature words and their contexts. The experimental results show that KDSL outperforms several representative state-of-the-art methods on various major benchmarks. Interestingly, it performs relatively well even when manually labeled data is unavailable, thus provides a potential solution for similar tasks in a lack of manual annotations.

Recursive neural networks have widely been used by researchers to handle applications with recursively or hierarchically structured data. However, embedded control flow deep learning frameworks such as TensorFlow, Theano, Caffe2, and MXNet fail to efficiently represent and execute such neural networks, due to lack of support for recursion. In this paper, we add recursion to the programming model of existing frameworks by complementing their design with recursive execution of dataflow graphs as well as additional APIs for recursive definitions. Unlike iterative implementations, which can only understand the topological index of each node in recursive data structures, our recursive implementation is able to exploit the recursive relationships between nodes for efficient execution based on parallel computation. We present an implementation on TensorFlow and evaluation results with various recursive neural network models, showing that our recursive implementation not only conveys the recursive nature of recursive neural networks better than other implementations, but also uses given resources more effectively to reduce training and inference time.

Simply put, the reason is that classifiers are Quantification can in principle be solved by classifying all the unlabelled typically trained to minimize classification error, which is by and items and counting how many of them have been attributed large proportional to (FP FN), while a good quantifier should to each class. However, this "classify and count" approach has been be trained to minimise quantification error, which is by and large shown to yield suboptimal quantification accuracy; this has established proportional to FP FN (where TP, FP, FN, T N denote the usual quantification as a task of its own, and given rise to a number counts from a binary contingency table). of methods specifically devised for it. We propose a recurrent neural In this paper we tackle quantification in a binary setting, and propose network architecture for quantification (that we call QuaNet) that a recurrent neural network architecture (that we call QuaNet) observes the classification predictions to learn higher-order "quantification that observes the classification predictions to learn higher-order embeddings", which are then refined by incorporating "quantification embeddings", which are then refined by incorporating quantification predictions of simple classify-and-count-like methods.

In this paper, we propose a new kernel-based co-occurrence measure that can be applied to sparse linguistic expressions (e.g., sentences) with a very short learning time, as an alternative to pointwise mutual information (PMI). As well as deriving PMI from mutual information, we derive this new measure from the Hilbert--Schmidt independence criterion (HSIC); thus, we call the new measure the pointwise HSIC (PHSIC). PHSIC can be interpreted as a smoothed variant of PMI that allows various similarity metrics (e.g., sentence embeddings) to be plugged in as kernels. Moreover, PHSIC can be estimated by simple and fast (linear in the size of the data) matrix calculations regardless of whether we use linear or nonlinear kernels. Empirically, in a dialogue response selection task, PHSIC is learned thousands of times faster than an RNN-based PMI while outperforming PMI in accuracy. In addition, we also demonstrate that PHSIC is beneficial as a criterion of a data selection task for machine translation owing to its ability to give high (low) scores to a consistent (inconsistent) pair with other pairs.

Robust risk minimisation has several advantages: it has been studied with regards to improving the generalisation properties of models and robustness to adversarial perturbation. We bound the distributionally robust risk for a model class rich enough to include deep neural networks by a regularised empirical risk involving the Lipschitz constant of the model. This allows us to interpret and quantify the robustness properties of a deep neural network. As an application we show the distributionally robust risk upperbounds the adversarial training risk.

We introduce a hierarchical architecture for video understanding that exploits the structure of real world actions by capturing targets at different levels of granularity. We design the model such that it first learns simpler coarse-grained tasks, and then moves on to learn more fine-grained targets. The model is trained with a joint loss on different granularity levels. We demonstrate empirical results on the recent release of Something-Something dataset, which provides a hierarchy of targets, namely coarse-grained action groups, fine-grained action categories, and captions. Experiments suggest that models that exploit targets at different levels of granularity achieve better performance on all levels.