A wide variety of DNN applications increasingly rely on the cloud to perform their huge computation. This heavy trend toward cloud-hosted inference services raises serious privacy concerns. This model requires the sending of private and privileged data over the network to remote servers, exposing it to the service provider. Even if the provider is trusted, the data can still be vulnerable over communication channels or via side-channel attacks [1,2] at the provider. To that end, this paper aims to reduce the information content of the communicated data without compromising the cloud service's ability to provide a DNN inference with acceptably high accuracy. This paper presents an end-to-end framework, called Shredder, that, without altering the topology or the weights of a pre-trained network, learns an additive noise distribution that significantly reduces the information content of communicated data while maintaining the inference accuracy. Shredder learns the additive noise by casting it as a tensor of trainable parameters enabling us to devise a loss functions that strikes a balance between accuracy and information degradation. The loss function exposes a knob for a disciplined and controlled asymmetric trade-off between privacy and accuracy. While keeping the DNN intact, Shredder enables inference on noisy data without the need to update the model or the cloud. Experimentation with real-world DNNs shows that Shredder reduces the mutual information between the input and the communicated data to the cloud by 70.2% compared to the original execution while only sacrificing 1.46% loss in accuracy.

Recent breakthroughs in the field of deep learning have led to advancements in a broad spectrum of tasks in computer vision, audio processing, natural language processing and other areas. In most instances where these tasks are deployed in real-world scenarios, the models used in them have been shown to be susceptible to adversarial attacks, making it imperative for us to address the challenge of their adversarial robustness. Existing techniques for adversarial robustness fall into three broad categories: defensive distillation techniques, adversarial training techniques, and randomized or non-deterministic model based techniques. In this paper, we propose a novel neural network paradigm that falls under the category of randomized models for adversarial robustness, but differs from all existing techniques under this category in that it models each parameter of the network as a statistical distribution with learnable parameters. We show experimentally that this framework is highly robust to a variety of white-box and black-box adversarial attacks, while preserving the task-specific performance of the traditional neural network model.

Different from the traditional classification tasks which assume mutual exclusion of labels, hierarchical multi-label classification (HMLC) aims to assign multiple labels to every instance with the labels organized under hierarchical relations. In fact, linguistic ontologies are intrinsic hierarchies. Besides the labels, the conceptual relations between words can also form hierarchical structures. Thus it can be a challenge to learn mappings from the word space to the label space, and vice versa. We propose to model the word and label hierarchies by embedding them jointly in the hyperbolic space. The main reason is that the tree-likeness of the hyperbolic space matches the complexity of symbolic data with hierarchical structures. A new hyperbolic interaction model (HyperIM) is designed to learn the label-aware document representations and make predictions for HMLC. Extensive experiments are conducted on three benchmark datasets. The results have demonstrated that the new model can realistically capture the complex data structures and further improve the performance for HMLC comparing with the state-of-the-art methods. To facilitate future research, our code is publicly available.

In this work we address the problem of argument search. The purpose of argument search is the distillation of pro and contra arguments for requested topics from large text corpora. In previous works, the usual approach is to use a standard search engine to extract text parts which are relevant to the given topic and subsequently use an argument recognition algorithm to select arguments from them. The main challenge in the argument recognition task, which is also known as argument mining, is that often sentences containing arguments are structurally similar to purely informative sentences without any stance about the topic. In fact, they only differ semantically. Most approaches use topic or search term information only for the first search step and therefore assume that arguments can be classified independently of a topic. We argue that topic information is crucial for argument mining, since the topic defines the semantic context of an argument. Precisely, we propose different models for the classification of arguments, which take information about a topic of an argument into account. Moreover, to enrich the context of a topic and to let models understand the context of the potential argument better, we integrate information from different external sources such as Knowledge Graphs or pre-trained NLP models. Our evaluation shows that considering topic information, especially in connection with external information, provides a significant performance boost for the argument mining task.

Image captioning applied to biomedical images can assist and accelerate the diagnosis process followed by clinicians. This article is the first survey of biomedical image captioning, discussing datasets, evaluation measures, and state of the art methods. Additionally, we suggest two baselines, a weak and a stronger one; the latter outperforms all current state of the art systems on one of the datasets.

We propose interpretable deep Gaussian Processes (GPs) that combine the expressiveness of deep Neural Networks (NNs) with quantified uncertainty of deep GPs. Our approach is based on approximating deep GP as a GP, which allows explicit, analytic forms for compositions of a wide variety of kernels. Consequently, our approach admits interpretation as both NNs with specified activation functions and as a variational approximation to deep GPs. We provide general recipes for deriving the effective kernels for deep GPs of two, three, or infinitely many layers, composed of homogeneous or heterogeneous kernels. Results illustrate the expressiveness of our effective kernels through samples from the prior and inference on simulated data and demonstrate advantages of interpretability by analysis of analytic forms, drawing relations and equivalences across kernels, and a priori identification of non-pathological regimes of hyperparameter space.

Graph Convolutional Neural Networks (graph CNNs) are a promising deep learning approach for analyzing graph-structured data. However, it is known that they do not improve (or sometimes worsen) their predictive performance as we pile up more layers and make them deeper. To tackle this problem, we investigate the expressive power of graph CNNs by analyzing their asymptotic behaviors as the layer size tends to infinity. Our strategy is to generalize the forward propagation of a Graph Convolutional Network (GCN), which is one of the most popular graph CNN variants, as a specific dynamical system. In the case of GCNs, we show that when the weights satisfy the conditions determined by the spectra of the (augmented) normalized Laplacian, the output of GCNs exponentially approaches the set of signals that carry only information of the connected components and node degrees for distinguishing nodes. Our theory enables us to directly relate the expressive power of GCNs with the topological information of the underlying graphs, which is inherent in the graph spectra. To demonstrate this, we characterize the asymptotic behavior of GCNs on the Erd\H{o}s -- R\'{e}nyi graph. We show that when the Erd\H{o}s -- R\'{e}nyi graph is sufficiently dense and large, a wide range of GCNs on them suffers from this ``information loss" in the limit of infinite layers with high probability. Furthermore, our theory provides principled guidelines for the weight normalization of graph CNNs. We experimentally confirmed that weight scaling based on our theory enhanced the predictive performance of GCNs in real data.

A new variational autoencoder (VAE) model is proposed that learns a succinct common representation of two correlated data variables for conditional and joint generation tasks. The proposed Wyner VAE model is based on two information theoretic problems---distributed simulation and channel synthesis---in which Wyner's common information arises as the fundamental limit of the succinctness of the common representation. The Wyner VAE decomposes a pair of correlated data variables into their common representation (e.g., a shared concept) and local representations that capture the remaining randomness (e.g., texture and style) in respective data variables by imposing the mutual information between the data variables and the common representation as a regularization term. The utility of the proposed approach is demonstrated through experiments for joint and conditional generation with and without style control using synthetic data and real images. Experimental results show that learning a succinct common representation achieves better generative performance and that the proposed model outperforms existing VAE variants and the variational information bottleneck method.

Residual Neural Networks (ResNets) achieve state-of-the-art performance in many computer vision problems. Compared to plain networks without residual connections (PlnNets), ResNets train faster, generalize better, and suffer less from the so-called degradation problem. We introduce simplified (but still nonlinear) versions of ResNets and PlnNets for which these discrepancies still hold, although to a lesser degree. We establish a 1-1 mapping between simplified ResNets and simplified PlnNets, and show that they are exactly equivalent to each other in expressive power for the same computational complexity. We conjecture that ResNets generalize better because they have better noise stability, and empirically support it for both simplified and fully-fledged networks.

Communication overhead is a major bottleneck hampering the scalability of distributed machine learning systems. Recently, there has been a surge of interest in using gradient compression to improve the communication efficiency of distributed neural network training. Using 1-bit quantization, signSGD with majority vote achieves a 32x reduction on communication cost. However, its convergence is based on unrealistic assumptions and can diverge in practice. In this paper, we propose a general distributed compressed SGD with Nesterov's momentum. We consider two-way compression, which compresses the gradients both to and from workers. Convergence analysis on nonconvex problems for general gradient compressors is provided. By partitioning the gradient into blocks, a blockwise compressor is introduced such that each gradient block is compressed and transmitted in 1-bit format with a scaling factor, leading to a nearly 32x reduction on communication. Experimental results show that the proposed method converges as fast as full-precision distributed momentum SGD and achieves the same testing accuracy. In particular, on distributed ResNet training with 7 workers on the ImageNet, the proposed algorithm achieves the same testing accuracy as momentum SGD using full-precision gradients, but with $46\%$ less wall clock time.