The high model complexity of deep learning algorithms enables remarkable learning capacity in many application domains. However, a large number of trainable parameters comes with a high cost. For example, during both the training and inference phases, the numerous trainable parameters consume a large amount of resources, such as CPU/GPU cores, memory and electric power. In addition, from a theoretical statistical learning perspective, the high complexity of the network can result in a high variance in its generalization performance. One way to reduce the complexity of a network without sacrificing its accuracy is to define and identify redundancies in order to remove them. In this work, we propose a method to observe and analyze redundancies in the weights of a 2D convolutional (Conv2D) network. Based on the proposed analysis, we construct a new layer called Separable Convolutional Eigen-Filters (SCEF) as an alternative parameterization to Conv2D layers. A SCEF layer can be easily implemented using depthwise separable convolution, which are known to be computationally effective. To verify our hypothesis, experiments are conducted on the CIFAR-10 and ImageNet datasets by replacing the Conv2D layers with SCEF and the results have shown an increased accuracy using about 2/3 of the original parameters and reduce the number of FLOPs to 2/3 of the original net. Implementation-wise, our method is highly modular, easy to use, fast to process and does not require any additional dependencies.

Learning workable representations of dynamical systems is becoming an increasingly important problem in a number of application areas. By leveraging recent work connecting deep neural networks to systems of differential equations, we propose variational integrator networks, a class of neural network architectures designed to ensure faithful representations of the dynamics under study. This class of network architectures facilitates accurate long-term prediction, interpretability, and data-efficient learning, while still remaining highly flexible and capable of modeling complex behavior. We demonstrate that they can accurately learn dynamical systems from both noisy observations in phase space and from image pixels within which the unknown dynamics are embedded.

Adversarial testing methods based on Projected Gradient Descent (PGD) are widely used for searching norm-bounded perturbations that cause the inputs of neural networks to be misclassified. This paper takes a deeper look at these methods and explains the effect of different hyperparameters (i.e., optimizer, step size and surrogate loss). We introduce the concept of MultiTargeted testing, which makes clever use of alternative surrogate losses, and explain when and how MultiTargeted is guaranteed to find optimal perturbations. Finally, we demonstrate that MultiTargeted outperforms more sophisticated methods and often requires less iterative steps than other variants of PGD found in the literature. Notably, MultiTargeted ranks first on MadryLab's white-box MNIST and CIFAR-10 leaderboards, reducing the accuracy of their MNIST model to 88.36% (with $\ell_\infty$ perturbations of $\epsilon = 0.3$) and the accuracy of their CIFAR-10 model to 44.03% (at $\epsilon = 8/255$). MultiTargeted also ranks first on the TRADES leaderboard reducing the accuracy of their CIFAR-10 model to 53.07% (with $\ell_\infty$ perturbations of $\epsilon = 0.031$).

We adapt the optimization's concept of momentum to reinforcement learning. Seeing the state-action value functions as an analog to the gradients in optimization, we interpret momentum as an average of consecutive $q$-functions. We derive Momentum Value Iteration (MoVI), a variation of Value Iteration that incorporates this momentum idea. Our analysis shows that this allows MoVI to average errors over successive iterations. We show that the proposed approach can be readily extended to deep learning. Specifically, we propose a simple improvement on DQN based on MoVI, and experiment it on Atari games.

Deep neural networks with convolutional layers usually process the entire spectrogram of an audio signal with the same time-frequency resolutions, number of filters, and dimensionality reduction scale. According to the constant-Q transform, good features can be extracted from audio signals if the low frequency bands are processed with high frequency resolution filters and the high frequency bands with high time resolution filters. In the spectrogram of a mixture of singing voices and music signals, there is usually more information about the voice in the low frequency bands than the high frequency bands. These raise the need for processing each part of the spectrogram differently. In this paper, we propose a multi-band multi-resolution fully convolutional neural network (MBR-FCN) for singing voice separation. The MBR-FCN processes the frequency bands that have more information about the target signals with more filters and smaller dimentionality reduction scale than the bands with less information. Furthermore, the MBR-FCN processes the low frequency bands with high frequency resolution filters and the high frequency bands with high time resolution filters. Our experimental results show that the proposed MBR-FCN with very few parameters achieves better singing voice separation performance than other deep neural networks.

Biomedical text tagging systems are plagued by the dearth of labeled training data. There have been recent attempts at using pre-trained encoders to deal with this issue. Pre-trained encoder provides representation of the input text which is then fed to task-specific layers for classification. The entire network is fine-tuned on the labeled data from the target task. Unfortunately, a low-resource biomedical task often has too few labeled instances for satisfactory fine-tuning. Also, if the label space is large, it contains few or no labeled instances for majority of the labels. Most biomedical tagging systems treat labels as indexes, ignoring the fact that these labels are often concepts expressed in natural language e.g. `Appearance of lesion on brain imaging'. To address these issues, we propose constructing extra labeled instances using label-text (i.e. label's name) as input for the corresponding label-index (i.e. label's index). In fact, we propose a number of strategies for manufacturing multiple artificial labeled instances from a single label. The network is then fine-tuned on a combination of real and these newly constructed artificial labeled instances. We evaluate the proposed approach on an important low-resource biomedical task called \textit{PICO annotation}, which requires tagging raw text describing clinical trials with labels corresponding to different aspects of the trial i.e. PICO (Population, Intervention/Control, Outcome) characteristics of the trial. Our empirical results show that the proposed method achieves a new state-of-the-art performance for PICO annotation with very significant improvements over competitive baselines.

Deep convolutional neural networks have achieved great successes over recent years, particularly in the domain of computer vision. They are fast, convenient, and -- thanks to mature frameworks -- relatively easy to implement and deploy. However, their reasoning is hidden inside a black box, in spite of a number of proposed approaches that try to provide human-understandable explanations for the predictions of neural networks. It is still a matter of debate which of these explainers are best suited for which situations, and how to quantitatively evaluate and compare them. In this contribution, we focus on the capabilities of explainers for convolutional deep neural networks in an extreme situation: a setting in which humans and networks fundamentally disagree. Deep neural networks are susceptible to adversarial attacks that deliberately modify input samples to mislead a neural network's classification, without affecting how a human observer interprets the input. Our goal with this contribution is to evaluate explainers by investigating whether they can identify adversarially attacked regions of an image. In particular, we quantitatively and qualitatively investigate the capability of three popular explainers of classifications -- classic salience, guided backpropagation, and LIME -- with respect to their ability to identify regions of attack as the explanatory regions for the (incorrect) prediction in representative examples from image classification. We find that LIME outperforms the other explainers.

Spatio-temporal event data are becoming increasingly available in a wide variety of applications, such as electronic transaction records, social network data, and crime data. How to efficiently detect anomalies in these dynamic systems using these streaming event data? We propose a novel anomaly detection framework for such event data combining the Long short-term memory (LSTM) and marked spatio-temporal point processes. The detection procedure can be computed in an online and distributed fashion via feeding the streaming data through an LSTM and a neural network-based discriminator. We study the false-alarm-rate and detection delay using theory and simulation and show that it can achieve weak signal detection by aggregating local statistics over time and networks. Finally, we demonstrate the good performance using real-world datasets.

Recent studies on the adversarial vulnerability of neural networks have shown that models trained with the objective of minimizing an upper bound on the worst-case loss over all possible adversarial perturbations improve robustness against adversarial attacks. Beside exploiting adversarial training framework, we show that by enforcing a Deep Neural Network (DNN) to be linear in transformed input and feature space improves robustness significantly. We also demonstrate that by augmenting the objective function with Local Lipschitz regularizer boost robustness of the model further. Our method outperforms most sophisticated adversarial training methods and achieves state of the art adversarial accuracy on MNIST, CIFAR10 and SVHN dataset. In this paper, we also propose a novel adversarial image generation method by leveraging Inverse Representation Learning and Linearity aspect of an adversarially trained deep neural network classifier.

--T opological data analysis aims to extract topological quantities from data, which tend to focus on the broader global structure of the data rather than local information. The Mapper method, specifically, generalizes clustering methods to identify significant global mathematical structures, which are out of reach of many other approaches. We propose a classifier based on applying the Mapper algorithm to data projected onto a latent space. We obtain the latent space by using PCA or autoencoders. Notably, a classifier based on the Mapper method is immune to any gradient based attack, and improves robustness over traditional CNNs (convolutional neural networks). We report theoretical justification and some numerical experiments that confirm our claims. I NTRODUCTION Deep neural networks [1], [2] are well known to be not robust with respect to input image perturbations, which are designed by adding to images perturbations that are typically non-perceptible by humans [3]-[5]. In this paper we explore opportunities for combining deep learning techniques with a well known topological data analysis (TDA) algorithm - the Mapper algorithm [6], which we use to create classifiers with improved robustness. First, the training data is projected onto a latent space. The latent space in the simplest variant is constructed using PCA components, and we also use nonlinear projections by utilizing various autoencoders [1], [7]-[9]. Then, a discrete graph representation (Mapper) is assigned to the training data projected onto the latent space .