Instead of training individual networks with different width configurations, we train a shared network with switchable batch normalization. At runtime, the network can adjust its width on the fly according to on-device benchmarks and resource constraints, rather than downloading and offloading different models. Our trained networks, named slimmable neural networks, achieve similar (and in many cases better) ImageNet classification accuracy than individually trained models of MobileNet v1, MobileNet v2, ShuffleNet and ResNet-50 at different widths respectively. We also demonstrate better performance of slimmable models compared with individual ones across a wide range of applications including COCO bounding-box object detection, instance segmentation and person keypoint detection without tuning hyper-parameters. Lastly we visualize and discuss the learned features of slimmable networks. Recently deep neural networks are prevailing in applications on mobile phones, augmented reality devices and autonomous cars. Many of these applications require a short response time. Towards this goal, manually designed lightweight networks (Howard et al., 2017; Zhang et al., 2017; Sandler et al., 2018) are proposed with low computational complexities and small memory footprints. Automated neural architecture search methods (Tan et al., 2018) also integrate on-device latency into search objectives by running models on a specific phone.

The sharp and recent increase in the availability of data captured by different sensors combined with their considerably heterogeneous natures poses a serious challenge for the effective and efficient processing of remotely sensed data. Such an increase in remote sensing and ancillary datasets, however, opens up the possibility of utilizing multimodal datasets in a joint manner to further improve the performance of the processing approaches with respect to the application at hand. Multisource data fusion has, therefore, received enormous attention from researchers worldwide for a wide variety of applications. Moreover, thanks to the revisit capability of several spaceborne sensors, the integration of the temporal information with the spatial and/or spectral/backscattering information of the remotely sensed data is possible and helps to move from a representation of 2D/3D data to 4D data structures, where the time variable adds new information as well as challenges for the information extraction algorithms. There are a huge number of research works dedicated to multisource and multitemporal data fusion, but the methods for the fusion of different modalities have expanded in different paths according to each research community. This paper brings together the advances of multisource and multitemporal data fusion approaches with respect to different research communities and provides a thorough and discipline-specific starting point for researchers at different levels (i.e., students, researchers, and senior researchers) willing to conduct novel investigations on this challenging topic by supplying sufficient detail and references.

Given the recent success of Deep Learning applied to a variety of single tasks, it is natural to consider more human-realistic settings. Perhaps the most difficult of these settings is that of continual lifelong learning, where the model must learn online over a continuous stream of non-stationary data. A successful continual lifelong learning system must have three key capabilities: it must learn and adapt over time, it must not forget what it has learned, and it must be efficient in both training time and memory. Recent techniques have focused their efforts primarily on the first two capabilities while questions of efficiency remain largely unexplored. In this paper, we consider the problem of efficient and effective storage of experiences over very large time-frames. In particular we consider the case where typical experiences are O(n) bits and memories are limited to O(k) bits for k << n. We present a novel scalable architecture and training algorithm in this challenging domain and provide an extensive evaluation of its performance. Our results show that we can achieve considerable gains on top of state-of-the-art methods such as GEM.

Deep neural networks have achieved state-of-the-art accuracies in a wide range of computer vision, speech recognition, and machine translation tasks. However the limits of memory bandwidth and computational power constrain the range of devices capable of deploying these modern networks. To address this problem, we propose SQuantizer, a new training method that jointly optimizes for both sparse and low-precision neural networks while maintaining high accuracy and providing a high compression rate. This approach brings sparsification and low-bit quantization into a single training pass, employing these techniques in an order demonstrated to be optimal. Our method achieves state-of-the-art accuracies using 4-bit and 2-bit precision for ResNet18, MobileNet-v2 and ResNet50, even with high degree of sparsity. The compression rates of 18x for ResNet18 and 17x for ResNet50, and 9x for MobileNet-v2 are obtained when SQuantizing both weights and activations within 1% and 2% loss in accuracy for ResNets and MobileNet-v2 respectively. An extension of these techniques to object detection also demonstrates high accuracy on YOLO-v2. Additionally, our method allows for fast single pass training, which is important for rapid prototyping and neural architecture search techniques. Finally extensive results from this simultaneous training approach allows us to draw some useful insights into the relative merits of sparsity and quantization.

In this paper, we introduce a new task - interactive image editing via conversational language, where users can guide an agent to edit images via multi-turn dialogue in natural language. In each dialogue turn, the agent takes a source image and a natural language description from the user as the input, and generates a target image following the textual description. Two new datasets are created for this task,Zap-Seq and DeepFashion-Seq, collected via crowdsourcing. For this task, we propose a new Sequential Attention Genrative Adversarial Network (SeqAttnGAN) framework, which applies a neural state tracker to encode both source image and textual descriptions, and generates high quality images in each dialogue turn. To achieve better region specific text-to-image generation, we also introducean attention mechanism into the model. Experiments on the two datasets, including quantitative evaluation and user study, show that our model outperforms state-of-the-art ap-proaches in both image quality and text-to-image consistency.

Research on time-series similarity measures has emphasized the need for elastic methods which align the indices of pairs of time series and a plethora of non-parametric have been proposed for the task. On the other hand, deep learning approaches are dominant in closely related domains, such as learning image and text sentence similarity. In this paper, we propose \textit{NeuralWarp}, a novel measure that models the alignment of time-series indices in a deep representation space, by modeling a warping function as an upper level neural network between deeply-encoded time series values. Experimental results demonstrate that \textit{NeuralWarp} outperforms both non-parametric and un-warped deep models on a range of diverse real-life datasets.

The length of the geodesic between two data points along the Riemannian manifold, induced by a deep generative model, yields a principled measure of similarity. Applications have so far been limited to low-dimensional latent spaces, as the method is computationally demanding: it constitutes to solving a non-convex optimisation problem. Our approach is to tackle a relaxation: finding shortest paths in a finite graph of samples from the aggregate approximate posterior can be solved exactly, at greatly reduced runtime, and without notable loss in quality. The method is hence applicable to high-dimensional problems in the visual domain. We validate the approach empirically on a series of experiments using variational autoencoders applied to image data, tackling the Chair, Faces and FashionMNIST data sets.

The retinal vascular condition is a reliable biomarker of several ophthalmologic and cardiovascular diseases, so automatic vessel segmentation may be crucial to diagnose and monitor them. In this paper, we propose a novel method that combines the multiscale analysis provided by the Stationary Wavelet Transform with a multiscale Fully Convolutional Neural Network to cope with the varying width and direction of the vessel structure in the retina. Our proposal uses rotation operations as the basis of a joint strategy for both data augmentation and prediction, which allows us to explore the information learned during training to refine the segmentation. The method was evaluated on three publicly available databases, achieving an average accuracy of 0.9576, 0.9694, and 0.9653, and average area under the ROC curve of 0.9821, 0.9905, and 0.9855 on the DRIVE, STARE, and CHASE_DB1 databases, respectively. It also appears to be robust to the training set and to the inter-rater variability, which shows its potential for real-world applications.

In this paper, we explore the utilization of natural language to drive transfer for reinforcement learning (RL). Despite the wide-spread application of deep RL techniques, learning generalized policy representations that work across domains remains a challenging problem. We demonstrate that textual descriptions of environments provide a compact intermediate channel to facilitate effective policy transfer. Specifically, by learning to ground the meaning of text to the dynamics of the environment such as transitions and rewards, an autonomous agent can effectively bootstrap policy learning on a new domain given its description. We employ a model-based RL approach consisting of a differentiable planning module, a model-free component and a factorized state representation to effectively use entity descriptions. Our model outperforms prior work on both transfer and multi-task scenarios in a variety of different environments. For instance, we achieve up to 14% and 11.5% absolute improvement over previously existing models in terms of average and initial rewards, respectively.

Multimodal sentiment analysis is a core research area that studies speaker sentiment expressed from the language, visual, and acoustic modalities. The central challenge in multimodal learning involves inferring joint representations that can process and relate information from these modalities. However, existing work learns joint representations by requiring all modalities as input and as a result, the learned representations may be sensitive to noisy or missing modalities at test time. With the recent success of sequence to sequence (Seq2Seq) models in machine translation, there is an opportunity to explore new ways of learning joint representations that may not require all input modalities at test time. In this paper, we propose a method to learn robust joint representations by translating between modalities. Our method is based on the key insight that translation from a source to a target modality provides a method of learning joint representations using only the source modality as input. We augment modality translations with a cycle consistency loss to ensure that our joint representations retain maximal information from all modalities. Once our translation model is trained with paired multimodal data, we only need data from the source modality at test time for final sentiment prediction. This ensures that our model remains robust from perturbations or missing information in the other modalities. We train our model with a coupled translation-prediction objective and it achieves new state-of-the-art results on multimodal sentiment analysis datasets: CMU-MOSI, ICT-MMMO, and YouTube. Additional experiments show that our model learns increasingly discriminative joint representations with more input modalities while maintaining robustness to missing or perturbed modalities.