Gradient descent yields zero training loss in polynomial time for deep neural networks despite non-convex nature of the objective function. The behavior of network in the infinite width limit trained by gradient descent can be described by the Neural Tangent Kernel (NTK) introduced in \cite{Jacot2018Neural}. In this paper, we study dynamics of the NTK for finite width Deep Residual Network (ResNet) using the neural tangent hierarchy (NTH) proposed in \cite{Huang2019Dynamics}. For a ResNet with smooth and Lipschitz activation function, we reduce the requirement on the layer width $m$ with respect to the number of training samples $n$ from quartic to cubic. Our analysis suggests strongly that the particular skip-connection structure of ResNet is the main reason for its triumph over fully-connected network.

Formation mechanisms are fundamental to the study of complex networks, but learning them from observations is challenging. In real-world domains, one often has access only to the final constructed graph, instead of the full construction process, and observed graphs exhibit complex structural properties. In this work, we propose GraphOpt, an end-to-end framework that jointly learns an implicit model of graph structure formation and discovers an underlying optimization mechanism in the form of a latent objective function. The learned objective can serve as an explanation for the observed graph properties, thereby lending itself to transfer across different graphs within a domain. GraphOpt poses link formation in graphs as a sequential decision-making process and solves it using maximum entropy inverse reinforcement learning algorithm. Further, it employs a novel continuous latent action space that aids scalability. Empirically, we demonstrate that GraphOpt discovers a latent objective transferable across graphs with different characteristics. GraphOpt also learns a robust stochastic policy that achieves competitive link prediction performance without being explicitly trained on this task and further enables construction of graphs with properties similar to those of the observed graph.

The rapid growth and distribution of IT systems increases their complexity and aggravates operation and maintenance. To sustain control over large sets of hosts and the connecting networks, monitoring solutions are employed and constantly enhanced. They collect diverse key performance indicators (KPIs) (e.g. CPU utilization, allocated memory, etc.) and provide detailed information about the system state. Storing such metrics over a period of time naturally raises the motivation of predicting future KPI progress based on past observations. Although, a variety of time series forecasting methods exist, forecasting the progress of IT system KPIs is very hard. First, KPI types like CPU utilization or allocated memory are very different and hard to be expressed by the same model. Second, system components are interconnected and constantly changing due to soft- or firmware updates and hardware modernization. Thus a frequent model retraining or fine-tuning must be expected. Therefore, we propose a lightweight solution for KPI series prediction based on historic observations. It consists of a weighted heterogeneous ensemble method composed of two models - a neural network and a mean predictor. As ensemble method a weighted summation is used, whereby a heuristic is employed to set the weights. The modelling approach is evaluated on the available FedCSIS 2020 challenge dataset and achieves an overall $R^2$ score of 0.10 on the preliminary 10% test data and 0.15 on the complete test data. We publish our code on the following github repository: https://github.com/citlab/fed_challenge

Network embedding, aiming to project a network into a low-dimensional space, is increasingly becoming a focus of network research. Semi-supervised network embedding takes advantage of labeled data, and has shown promising performance. However, existing semi-supervised methods would get unappealing results in the completely-imbalanced label setting where some classes have no labeled nodes at all. To alleviate this, we propose two novel semi-supervised network embedding methods. The first one is a shallow method named RSDNE. Specifically, to benefit from the completely-imbalanced labels, RSDNE guarantees both intra-class similarity and inter-class dissimilarity in an approximate way. The other method is RECT which is a new class of graph neural networks. Different from RSDNE, to benefit from the completely-imbalanced labels, RECT explores the class-semantic knowledge. This enables RECT to handle networks with node features and multi-label setting. Experimental results on several real-world datasets demonstrate the superiority of the proposed methods.

Metal artifact reduction (MAR) is one of the most important research topics in computed tomography (CT). With the advance of deep learning technology for image reconstruction,various deep learning methods have been also suggested for metal artifact removal, among which supervised learning methods are most popular. However, matched non-metal and metal image pairs are difficult to obtain in real CT acquisition. Recently, a promising unsupervised learning for MAR was proposed using feature disentanglement, but the resulting network architecture is complication and difficult to handle large size clinical images. To address this, here we propose a much simpler and much effective unsupervised MAR method for CT. The proposed method is based on a novel beta-cycleGAN architecture derived from the optimal transport theory for appropriate feature space disentanglement. Another important contribution is to show that attention mechanism is the key element to effectively remove the metal artifacts. Specifically, by adding the convolutional block attention module (CBAM) layers with a proper disentanglement parameter, experimental results confirm that we can get more improved MAR that preserves the detailed texture of the original image.

Deep learning achieves a high generalization performance in practice, despite the non-convexity of the gradient descent learning problem. Recently, the inductive bias in deep learning has been studied through the characterization of local minima. In this study, we show that the distribution of parameters learned by gradient descent converges to a spectrum of the ridgelet transform based on a ridgelet analysis, which is a wavelet-like analysis developed for neural networks. This convergence is stronger than those shown in previous results, and guarantees the shape of the parameter distribution has been identified with the ridgelet spectrum. In numerical experiments with finite models, we visually confirm the resemblance between the distribution of learned parameters and the ridgelet spectrum. Our study provides a better understanding of the theoretical background of an inductive bias theory based on lazy regimes.

Do Transformers Need Deep Long-Range Memory? Daniluk et al. (2017) observed that an LSTM We do this with a selective Transformer-XL (Dai et al., 2019), a Transformer memorisation process; most of the finer details of variant specialised for long-range sequence modelling the text are quickly forgotten and we retain a relatively via the introduction of a cache of past activations, compact representation of the book's details. obtained state-of-the-art results in the four Early models of natural language used recurrent major LM benchmarks -- PTB (Mikolov et al., neural networks (RNNs) such as the Long Short-2010), LM1B (Chelba et al., 2013), Enwik8 (Hutter, Term Memory (Hochreiter and Schmidhuber, 1997) 2012), and WikiText (Merity et al., 2016).

The paper proposes a novel technique for representing templates and instances of concept classes. A template representation refers to the generic representation that captures the characteristics of an entire class. The proposed technique uses end-to-end deep learning to learn structured and composable representations from input images and discrete labels. The obtained representations are based on distance estimates between the distributions given by the class label and those given by contextual information, which are modeled as environments. We prove that the representations have a clear structure allowing to decompose the representation into factors that represent classes and environments. We evaluate our novel technique on classification and retrieval tasks involving different modalities (visual and language data).

Several machine learning applications involve the optimization of higher-order derivatives (e.g., gradients of gradients) during training, which can be expensive with respect to memory and computation even with automatic differentiation. As a typical example in generative modeling, score matching (SM) involves the optimization of the trace of a Hessian. To improve computing efficiency, we rewrite the SM objective and its variants in terms of directional derivatives, and present a generic strategy to efficiently approximate any-order directional derivative with finite difference (FD). Our approximation only involves function evaluations, which can be executed in parallel, and no gradient computations. Thus, it reduces the total computational cost while also improving numerical stability. We provide two instantiations by reformulating variants of SM objectives into the FD forms. Empirically, we demonstrate that our methods produce results comparable to the gradient-based counterparts while being much more computationally efficient.

Deep neural networks (DNNs) are known for their high prediction performance, especially in perceptual tasks such as object recognition or autonomous driving. Still, DNNs are prone to yield unreliable predictions when encountering completely new situations without indicating their uncertainty. Bayesian variants of DNNs (BDNNs), such as MC dropout BDNNs, do provide uncertainty measures. However, BDNNs are slow during test time because they rely on a sampling approach. Here we present a single shot MC dropout approximation that preserves the advantages of BDNNs without being slower than a DNN. Our approach is to analytically approximate for each layer in a fully connected network the expected value and the variance of the MC dropout signal. We evaluate our approach on different benchmark datasets and a simulated toy example. We demonstrate that our single shot MC dropout approximation resembles the point estimate and the uncertainty estimate of the predictive distribution that is achieved with an MC approach, while being fast enough for real-time deployments of BDNNs.