Deep Learning methods, specifically convolutional neural networks (CNNs), have seen a lot of success in the domain of image-based data, where the data offers a clearly structured topology in the regular lattice of pixels. This 4-neighbourhood topological simplicity makes the application of convolutional masks straightforward for time series data, such as video applications, but many high-dimensional time series data are not organised in regular lattices, and instead values may have adjacency relationships with non-trivial topologies, such as small-world networks or trees. In our application case, human kinematics, it is currently unclear how to generalise convolutional kernels in a principled manner. Therefore we define and implement here a framework for general graph-structured CNNs for time series analysis. Our algorithm automatically builds convolutional layers using the specified adjacency matrix of the data dimensions and convolutional masks that scale with the hop distance. In the limit of a lattice-topology our method produces the well-known image convolutional masks. We test our method first on synthetic data of arbitrarily-connected graphs and human hand motion capture data, where the hand is represented by a tree capturing the mechanical dependencies of the joints. We are able to demonstrate, amongst other things, that inclusion of the graph structure of the data dimensions improves model prediction significantly, when compared against a benchmark CNN model with only time convolution layers.

Deep neural networks are typically trained by optimizing a loss function with an SGD variant, in conjunction with a decaying learning rate, until convergence. We show that simple averaging of multiple points along the trajectory of SGD, with a cyclical or constant learning rate, leads to better generalization than conventional training. We also show that this Stochastic Weight Averaging (SWA) procedure finds much broader optima than SGD, and approximates the recent Fast Geometric Ensembling (FGE) approach with a single model. Using SWA we achieve notable improvement in test accuracy over conventional SGD training on a range of state-of-the-art residual networks, PyramidNets, DenseNets, and Shake-Shake networks on CIFAR-10, CIFAR-100, and ImageNet. In short, SWA is extremely easy to implement, improves generalization, and has almost no computational overhead.

Large-scale deep neural networks are both memory intensive and computation-intensive, thereby posing stringent requirements on the computing platforms. Hardware accelerations of deep neural networks have been extensively investigated in both industry and academia. Specific forms of binary neural networks (BNNs) and stochastic computing based neural networks (SCNNs) are particularly appealing to hardware implementations since they can be implemented almost entirely with binary operations. Despite the obvious advantages in hardware implementation, these approximate computing techniques are questioned by researchers in terms of accuracy and universal applicability. Also it is important to understand the relative pros and cons of SCNNs and BNNs in theory and in actual hardware implementations. In order to address these concerns, in this paper we prove that the "ideal" SCNNs and BNNs satisfy the universal approximation property with probability 1 (due to the stochastic behavior). The proof is conducted by first proving the property for SCNNs from the strong law of large numbers, and then using SCNNs as a "bridge" to prove for BNNs. Based on the universal approximation property, we further prove that SCNNs and BNNs exhibit the same energy complexity. In other words, they have the same asymptotic energy consumption with the growing of network size. We also provide a detailed analysis of the pros and cons of SCNNs and BNNs for hardware implementations and conclude that SCNNs are more suitable for hardware.

The availability of a large amount of electronic health records (EHR) provides huge opportunities to improve health care service by mining these data. One important application is clinical endpoint prediction, which aims to predict whether a disease, a symptom or an abnormal lab test will happen in the future according to patients' history records. This paper develops deep learning techniques for clinical endpoint prediction, which are effective in many practical applications. However, the problem is very challenging since patients' history records contain multiple heterogeneous temporal events such as lab tests, diagnosis, and drug administrations. The visiting patterns of different types of events vary significantly, and there exist complex nonlinear relationships between different events. In this paper, we propose a novel model for learning the joint representation of heterogeneous temporal events. The model adds a new gate to control the visiting rates of different events which effectively models the irregular patterns of different events and their nonlinear correlations. Experiment results with real-world clinical data on the tasks of predicting death and abnormal lab tests prove the effectiveness of our proposed approach over competitive baselines.

In industrial machine learning pipelines, data often arrive in parts. Particularly in the case of deep neural networks, it may be too expensive to train the model from scratch each time, so one would rather use a previously learned model and the new data to improve performance. However, deep neural networks are prone to getting stuck in a suboptimal solution when trained on only new data as compared to the full dataset. Our work focuses on a continuous learning setup where the task is always the same and new parts of data arrive sequentially. We apply a Bayesian approach to update the posterior approximation with each new piece of data and find this method to outperform the traditional approach in our experiments.

Binarization of digital documents is the task of classifying each pixel in an image of the document as belonging to the background (parchment/paper) or foreground (text/ink). Historical documents are often subject to degradations, that make the task challenging. In the current work a deep neural network architecture is proposed that combines a fully convolutional network with an unrolled primal-dual network that can be trained end-to-end in order to achieve state of the art binarization on four out of seven datasets. Document binarization is formulated as a energy minimization problem. A fully convolutional neural network is trained for semantic labeling of pixels to provide class labeling cost associated with each pixel. This cost estimate is refined along the edges to compensate for any over or under estimation of the under represented fore-ground class using a primal-dual approach. We provide necessary overview on proximal operator that facilitates theoretical underpinning in order to train a primal-dual network using a gradient descent algorithm. Numerical instabilities encountered due to the recurrent nature of primal-dual approach are handled. We provide experimental results on document binarization competition dataset along with network changes and hyperparameter tuning required for stability and performance of the network. The network when pre-trained on synthetic dataset performs better as per the competition metrics.

Computed Tomography (CT) reconstruction is a fundamental component to a wide variety of applications ranging from security, to healthcare. The classical techniques require measuring projections, called sinograms, from a full 180$^\circ$ view of the object. This is impractical in a limited angle scenario, when the viewing angle is less than 180$^\circ$, which can occur due to different factors including restrictions on scanning time, limited flexibility of scanner rotation, etc. The sinograms obtained as a result, cause existing techniques to produce highly artifact-laden reconstructions. In this paper, we propose to address this problem through implicit sinogram completion, on a challenging real world dataset containing scans of common checked-in luggage. We propose a system, consisting of 1D and 2D convolutional neural networks, that operates on a limited angle sinogram to directly produce the best estimate of a reconstruction. Next, we use the x-ray transform on this reconstruction to obtain a "completed" sinogram, as if it came from a full 180$^\circ$ measurement. We feed this to standard analytical and iterative reconstruction techniques to obtain the final reconstruction. We show with extensive experimentation that this combined strategy outperforms many competitive baselines. We also propose a measure of confidence for the reconstruction that enables a practitioner to gauge the reliability of a prediction made by our network. We show that this measure is a strong indicator of quality as measured by the PSNR, while not requiring ground truth at test time. Finally, using a segmentation experiment, we show that our reconstruction preserves the 3D structure of objects effectively.

Conditional Generative Adversarial Networks (cGANs) are generative models that can produce data samples ($x$) conditioned on both latent variables ($z$) and known auxiliary information ($c$). We propose the Bidirectional cGAN (BiCoGAN), which effectively disentangles $z$ and $c$ in the generation process and provides an encoder that learns inverse mappings from $x$ to both $z$ and $c$, trained jointly with the generator and the discriminator. We present crucial techniques for training BiCoGANs, which involve an extrinsic factor loss along with an associated dynamically-tuned importance weight. As compared to other encoder-based cGANs, BiCoGANs encode $c$ more accurately, and utilize $z$ and $c$ more effectively and in a more disentangled way to generate samples.

We propose a simple extension to the ReLU-family of activation functions that allows them to shift the mean activation across a layer towards zero. Combined with proper weight initialization, this alleviates the need for normalization layers. We explore the training of deep vanilla recurrent neural networks (RNNs) with up to 144 layers, and show that bipolar activation functions help learning in this setting. On the Penn Treebank and Text8 language modeling tasks we obtain competitive results, improving on the best reported results for non-gated networks. In experiments with convolutional neural networks without batch normalization, we find that bipolar activations produce a faster drop in training error, and results in a lower test error on the CIFAR-10 classification task.

Models applied on real time response task, like click-through rate (CTR) prediction model, require high accuracy and rigorous response time. Therefore, top-performing deep models of high depth and complexity are not well suited for these applications with the limitations on the inference time. In order to further improve the neural networks' performance given the time and computational limitations, we propose an approach that exploits a cumbersome net to help train the lightweight net for prediction. We dub the whole process rocket launching, where the cumbersome booster net is used to guide the learning of the target light net throughout the whole training process. We analyze different loss functions aiming at pushing the light net to behave similarly to the booster net, and adopt the loss with best performance in our experiments. We use one technique called gradient block to improve the performance of the light net and booster net further. Experiments on benchmark datasets and real-life industrial advertisement data present that our light model can get performance only previously achievable with more complex models.