Industrial forecasting has entered an era of unprecedented growth in the size and complexity of data which require new modeling methodologies. While many new general purpose machine learning approaches have emerged, they remain poorly understand and irreconcilable with more traditional statistical modeling approaches. We present a general class of exponential smoothed recurrent neural networks (RNNs) which are well suited to modeling non-stationary dynamical systems arising in industrial applications such as electricity load management and financial risk and trading. In particular, we analyze their capacity to characterize the non-linear partial autocorrelation structure of time series and directly capture dynamic effects such as seasonality and regime changes. Application of exponentially smoothed RNNs to electricity load forecasting, weather data and financial time series, such as minute level Bitcoin prices and CME futures tick data, highlight the efficacy of exponential smoothing for multi-step time series forecasting. The results also suggest that popular, but more complicated neural network architectures originally designed for speech processing, such as LSTMs and GRUs, are likely over-engineered for industrial forecasting and light-weight exponentially smoothed architectures capture the salient features while being superior and more robust than simple RNNs.

The inductive bias of a neural network is largely determined by the architecture and the training algorithm. To achieve good generalization, how to effectively train a neural network is even more important than designing the architecture. We propose a novel orthogonal over-parameterized training (OPT) framework that can provably minimize the hyperspherical energy which characterizes the diversity of neurons on a hypersphere. By constantly maintaining the minimum hyperspherical energy during training, OPT can greatly improve the network generalization. Specifically, OPT fixes the randomly initialized weights of the neurons and learns an orthogonal transformation that applies to these neurons. We propose multiple ways to learn such an orthogonal transformation, including unrolling orthogonalization algorithms, applying orthogonal parameterization, and designing orthogonality-preserving gradient update. Interestingly, OPT reveals that learning a proper coordinate system for neurons is crucial to generalization and may be more important than learning a specific relative position of neurons. We further provide theoretical insights of why OPT yields better generalization. Extensive experiments validate the superiority of OPT.

With the growth of interest in the attack and defense of deep neural networks, researchers are focusing more on the robustness of applying them to devices with limited memory. Thus, unlike adversarial training, which only considers the balance between accuracy and robustness, we come to a more meaningful and critical issue, i.e., the balance among accuracy, efficiency and robustness (AER). Recently, some related works focused on this issue, but with different observations, and the relations among AER remain unclear. This paper first investigates the robustness of pruned models with different compression ratios under the gradual pruning process and concludes that the robustness of the pruned model drastically varies with different pruning processes, especially in response to attacks with large strength. Second, we test the performance of mixing the clean data and adversarial examples (generated with a prescribed uniform budget) into the gradual pruning process, called adversarial pruning, and find the following: the pruned model's robustness exhibits high sensitivity to the budget. Furthermore, to better balance the AER, we propose an approach called blind adversarial pruning (BAP), which introduces the idea of blind adversarial training into the gradual pruning process. The main idea is to use a cutoff-scale strategy to adaptively estimate a nonuniform budget to modify the AEs used during pruning, thus ensuring that the strengths of AEs are dynamically located within a reasonable range at each pruning step and ultimately improving the overall AER of the pruned model. The experimental results obtained using BAP for pruning classification models based on several benchmarks demonstrate the competitive performance of this method: the robustness of the model pruned by BAP is more stable among varying pruning processes, and BAP exhibits better overall AER than adversarial pruning.

When it comes to preserving privacy in medical machine learning, two important considerations are (1) keeping data local to the institution and (2) avoiding inference of sensitive information from the trained model. These are often addressed using federated learning and differential privacy, respectively. However, the commonly used Federated Averaging algorithm requires a high degree of synchronization between participating institutions. For this reason, we turn our attention to Private Aggregation of Teacher Ensembles (PATE), where all local models can be trained independently without inter-institutional communication. The purpose of this paper is thus to explore how PATE -- originally designed for classification -- can best be adapted for semantic segmentation. To this end, we build low-dimensional representations of segmentation masks which the student can obtain through low-sensitivity queries to the private aggregator. On the Brain Tumor Segmentation (BraTS 2019) dataset, an Autoencoder-based PATE variant achieves a higher Dice coefficient for the same privacy guarantee than prior work based on noisy Federated Averaging.

New results suggest strong limits to the feasibility of classifying human brain activity evoked from image stimuli, as measured through EEG. Considerable prior work suffers from a confound between the stimulus class and the time since the start of the experiment. A prior attempt to avoid this confound using randomized trials was unable to achieve results above chance in a statistically significant fashion when the data sets were of the same size as the original experiments. Here, we again attempt to replicate these experiments with randomized trials on a far larger (20x) dataset of 1,000 stimulus presentations of each of forty classes, all from a single subject. To our knowledge, this is the largest such EEG data collection effort from a single subject and is at the bounds of feasibility. We obtain classification accuracy that is marginally above chance and above chance in a statistically significant fashion, and further assess how accuracy depends on the classifier used, the amount of training data used, and the number of classes. Reaching the limits of data collection without substantial improvement in classification accuracy suggests limits to the feasibility of this enterprise.

Adversarial training (AT) aims to improve the robustness of deep learning models by mixing clean data and adversarial examples (AEs). Most existing AT approaches can be grouped into restricted and unrestricted approaches. Restricted AT requires a prescribed uniform budget to constrain the magnitude of the AE perturbations during training, with the obtained results showing high sensitivity to the budget. On the other hand, unrestricted AT uses unconstrained AEs, resulting in the use of AEs located beyond the decision boundary; these overestimated AEs significantly lower the accuracy on clean data. These limitations mean that the existing AT approaches have difficulty in obtaining a comprehensively robust model with high accuracy and robustness when confronting attacks with varying strengths. Considering this problem, this paper proposes a novel AT approach named blind adversarial training (BAT) to better balance the accuracy and robustness. The main idea of this approach is to use a cutoff-scale strategy to adaptively estimate a nonuniform budget to modify the AEs used in the training, ensuring that the strengths of the AEs are dynamically located in a reasonable range and ultimately improving the overall robustness of the AT model. The experimental results obtained using BAT for training classification models on several benchmarks demonstrate the competitive performance of this method.

Generative adversarial networks (GANs) have attracted intense interest in the field of generative models. However, few investigations focusing either on the theoretical analysis or on algorithm design for the approximation ability of the generator of GANs have been reported. This paper will first theoretically analyze GANs' approximation property. Similar to the universal approximation property of the full connected neural networks with one hidden layer, we prove that the generator with the input latent variable in GANs can universally approximate the potential data distribution given the increasing hidden neurons. Furthermore, we propose an approach named stochastic data generation (SDG) to enhance GANs' approximation ability. Our approach is based on the simple idea of imposing randomness through data generation in GANs by a prior distribution on the conditional probability between the layers. Our approach can be easily implemented by using the reparameterization trick. The experimental results on synthetic dataset verify the improved approximation ability obtained by this SDG approach. In the practical dataset, the NSGAN/WGANGP with SDG can also outperform traditional GANs with little change in the model architectures.

Few-shot classification is a challenging task which aims to formulate the ability of humans to learn concepts from limited prior data and has drawn considerable attention in machine learning. Recent progress in few-shot classification has featured meta-learning, in which a parameterized model for a learning algorithm is defined and trained to learn the ability of handling classification tasks on extremely large or infinite episodes representing different classification task, each with a small labeled support set and its corresponding query set. In this work, we advance this few-shot classification paradigm by formulating it as a supervised classification learning problem. We further propose multi-episode and cross-way training techniques, which respectively correspond to the minibatch and pretraining in classification problems. Experimental results on a state-of-the-art few-shot classification method (prototypical networks) demonstrate that both the proposed training strategies can highly accelerate the training process without accuracy loss for varying few-shot classification problems on Omniglot and miniImageNet.

A generative adversarial network (GAN) based fully automatic electrocardiogram (ECG) arrhythmia classification system with high performance is presented in this paper. We have designed a discriminator (D) to take ECG coupling matrix as input, and then predict the input validity (real or generated) as well as arrhythmia classes. The generator (G) in our GAN is designed to generate various coupling matrix inputs conditioned on different arrhythmia classes for data augmentation. Upon completion of training for our GAN, we extracted the trained D as an arrhythmia classifier in a transfer learning manner. After fine-tuning D by including patient-specific normal beats estimated using an unsupervised algorithm, and generated abnormal beats by G that are usually rare to obtain, our fully automatic system showed superior overall classification performance for both supraventricular ectopic beats (SVEB or S beats) and ventricular ectopic beats (VEB or V beats) on the MIT-BIH arrhythmia database. It surpassed several state-of-art automatic classifiers and can perform on similar levels as some expert-assisted methods. In particular, the F1 score of SVEB has been improved by up to 12% over the top-performing automatic systems. Moreover, high sensitivity for both SVEB (85%) and VEB (93%) detection has been achieved, which is of great value for practical diagnosis. We, therefore, suggest our ACE-GAN (Generative Adversarial Network with Auxiliary Classifier for Electrocardiogram) based automatic system can be a promising and reliable tool for high throughput clinical screening practice, without any need of manual intervene or expert assisted labeling.

The transcriptomics of cancer tumors are characterized with tens of thousands of gene expression features. Patient prognosis or tumor stage can be assessed by machine learning techniques like supervised classification tasks given a gene expression profile. Feature selection is a useful approach to select the key genes which helps to classify tumors. In this work we propose a feature selection method based on Multiple Kernel Learning that results in a reduced subset of genes and a custom kernel that improves the classification performance when used in support vector classification. During the feature selection process this method performs a novel latent regularisation by relaxing the supervised target problem by introducing unsupervised structure obtained from the latent space learned by a non linear dimensionality reduction model. An improvement of the generalization capacity is obtained and assessed by the tumor classification performance on new unseen test samples when the classifier is trained with the features selected by the proposed method in comparison with other supervised feature selection approaches.