Deep convolutional neural networks are known to be unstable during training at high learning rate unless normalization techniques are employed. Normalizing weights or activations allows the use of higher learning rates, resulting in faster convergence and higher test accuracy. Batch normalization requires minibatch statistics that approximate the dataset statistics but this incurs additional compute and memory costs and causes a communication bottleneck for distributed training. Weight normalization and initialization-only schemes do not achieve comparable test accuracy. We introduce a new understanding of the cause of training instability and provide a technique that is independent of normalization and minibatch statistics. Our approach treats training instability as a spatial common mode signal which is suppressed by placing the model on a channel-wise zero-mean isocline that is maintained throughout training. Firstly, we apply channel-wise zero-mean initialization of filter kernels with overall unity kernel magnitude. At each training step we modify the gradients of spatial kernels so that their weighted channel-wise mean is subtracted in order to maintain the common mode rejection condition. This prevents the onset of mean shift. This new technique allows direct training of the test graph so that training and test models are identical. We also demonstrate that injecting random noise throughout the network during training improves generalization. This is based on the idea that, as a side effect, batch normalization performs deep data augmentation by injecting minibatch noise due to the weakness of the dataset approximation. Our technique achieves higher accuracy compared to batch normalization and for the first time shows that minibatches and normalization are unnecessary for state-of-the-art training.

ABSTRACT Speech emotion recognition is a challenging task and an important step towards more natural human-machine interaction. We show that pre-trained language models can be fine-tuned for text emotion recognition, achieving an accuracy of 69 .5 % on Task 4A of SemEval 2017, improving upon the previous state of the art by over 3 % absolute. We combine these language models with speech emotion recognition, achieving results of 73. 5 % accuracy when using provided transcriptions and speech data on a subset of four classes of the IEMOCAP dataset. For our experiments, we created IEmoNet, a modular and adaptable bimodal framework for speech emotion recognition based on pre-trained language models. Lastly, we discuss the idea of using an emotional classifier as a reward for reinforcement learning as a step towards more successful and convenient human-machine interaction. Index T erms-- Speech Emotion Recognition, Text Emotion Recognition, Bimodal Emotion Recognition, IEMOCAP, Self Attention, Pre-trained Language Models 1. INTRODUCTION Emotions are an important aspect of human behavior. They do not only influence the reaction to our environment [1, 2], but also actively change our perception of it [3] and sometimes even contribute to how well we remember specific events [4]. As such, they influence both human-human and human-machine interaction. However, in human-machine interaction, emotions are often not at all or only scarcely considered.

In recent years, large datasets of high-resolution mammalian neural images have become available, which has prompted active research on the analysis of gene expression data. Traditional image processing methods are typically applied for learning functional representations of genes, based on their expressions in these brain images. In this paper, we describe a novel end-to-end deep learning-based method for generating compact representations of in situ hybridization (ISH) images, which are invariant-to-translation. In contrast to traditional image processing methods, our method relies, instead, on deep convolutional denoising autoencoders (CDAE) for processing raw pixel inputs, and generating the desired compact image representations. We provide an in-depth description of our deep learning-based approach, and present extensive experimental results, demonstrating that representations extracted by CDAE can help learn features of functional gene ontology categories for their classification in a highly accurate manner. Our methods improve the previous state-of-the-art classification rate (Liscovitch, et al.) from an average AUC of 0.92 to 0.997, i.e., it achieves 96% reduction in error rate. Furthermore, the representation vectors generated due to our method are more compact in comparison to previous state-of-the-art methods, allowing for a more efficient high-level representation of images. These results are obtained with significantly downsampled images in comparison to the original high-resolution ones, further underscoring the robustness of our proposed method.

Malware allegedly developed by nation-states, also known as advanced persistent threats (APT), are becoming more common. The task of attributing an APT to a specific nation-state or classifying it to the correct APT family is challenging for several reasons. First, each nation-state has more than a single cyber unit that develops such malware, rendering traditional authorship attribution algorithms useless. Furthermore, the dataset of such available APTs is still extremely small. Finally, those APTs use state-of-the-art evasion techniques, making feature extraction challenging. In this paper, we use a deep neural network (DNN) as a classifier for nation-state APT attribution. We record the dynamic behavior of the APT when run in a sandbox and use it as raw input for the neural network, allowing the DNN to learn high level feature abstractions of the APTs itself. We also use the same raw features for APT family classification. Finally, we use the feature abstractions learned by the APT family classifier to solve the attribution problem. Using a test set of 1000 Chinese and Russian developed APTs, we achieved an accuracy rate of 98.6%.

We extend the Deep Galerkin Method (DGM) introduced in Sirignano and Spiliopoulos (2018) to solve a number of partial differential equations (PDEs) that arise in the context of optimal stochastic control and mean field games. First, we consider PDEs where the function is constrained to be positive and integrate to unity, as is the case with Fokker-Planck equations. Our approach involves reparameterizing the solution as the exponential of a neural network appropriately normalized to ensure both requirements are satisfied. This then gives rise to a partial integro-differential equation (PIDE) where the integral appearing in the equation is handled using importance sampling. Secondly, we tackle a number of Hamilton-Jacobi-Bellman (HJB) equations that appear in stochastic optimal control problems. The key contribution is that these equations are approached in their unsimplified primal form which includes an optimization problem as part of the equation. We extend the DGM algorithm to solve for the value function and the optimal control simultaneously by characterizing both as deep neural networks. Training the networks is performed by taking alternating stochastic gradient descent steps for the two functions, a technique similar in spirit to policy improvement algorithms.

Saliency maps are a popular approach to creating post-hoc explanations of image classifier outputs. These methods produce estimates of the relevance of each pixel to the classification output score, which can be displayed as a saliency map that highlights important pixels. Despite a proliferation of such methods, little effort has been made to quantify how good these saliency maps are at capturing the true relevance of the pixels to the classifier output (i.e. their "fidelity"). We therefore investigate existing metrics for evaluating the fidelity of saliency methods (i.e. saliency metrics). We find that there is little consistency in the literature in how such metrics are calculated, and show that such inconsistencies can have a significant effect on the measured fidelity. Further, we apply measures of reliability developed in the psychometric testing literature to assess the consistency of saliency metrics when applied to individual saliency maps. Our results show that saliency metrics can be statistically unreliable and inconsistent, indicating that comparative rankings between saliency methods generated using such metrics can be untrustworthy.

Recent advances in the sparse neural network literature have made it possible to prune many large feed forward and convolutional networks with only a small quantity of data. Yet, these same techniques often falter when applied to the problem of recovering sparse recurrent networks. These failures are quantitative: when pruned with recent techniques, RNNs typically obtain worse performance than they do under a simple random pruning scheme. The failures are also qualitative: the distribution of active weights in a pruned LSTM or GRU network tend to be concentrated in specific neurons and gates, and not well dispersed across the entire architecture. We seek to rectify both the quantitative and qualitative issues with recurrent network pruning by introducing a new recurrent pruning objective derived from the spectrum of the recurrent Jacobian. Our objective is data efficient (requiring only 64 data points to prune the network), easy to implement, and produces 95% sparse GRUs that significantly improve on existing baselines. We evaluate on sequential MNIST, Billion Words, and Wikitext.

The performance of deep neural networks is often attributed to their automated, task-related feature construction. It remains an open question, though, why this leads to solutions with good generalization, even in cases where the number of parameters is larger than the number of samples. Back in the 90s, Hochreiter and Schmidhuber observed that flatness of the loss surface around a local minimum correlates with low generalization error. For several flatness measures, this correlation has been empirically validated. However, it has recently been shown that existing measures of flatness cannot theoretically be related to generalization due to a lack of invariance with respect to reparameterizations. We propose a natural modification of existing flatness measures that results in invariance to reparameterization.

We propose the Square Attack, a new score-based black-box $l_2$ and $l_\infty$ adversarial attack that does not rely on local gradient information and thus is not affected by gradient masking. The Square Attack is based on a randomized search scheme where we select localized square-shaped updates at random positions so that the $l_\infty$- or $l_2$-norm of the perturbation is approximately equal to the maximal budget at each step. Our method is algorithmically transparent, robust to the choice of hyperparameters, and is significantly more query efficient compared to the more complex state-of-the-art methods. In particular, on ImageNet we improve the average query efficiency for various deep networks by a factor of at least $2$ and up to $7$ compared to the recent state-of-the-art $l_\infty$-attack of Meunier et al. while having a higher success rate. The Square Attack can even be competitive to gradient-based white-box attacks in terms of success rate. Moreover, we show its utility by breaking a recently proposed defense based on randomization. The code of our attack is available at https://github.com/max-andr/square-attack

Such behavior is desirable in multivariate structured prediction tasks, where handcrafted per-pixel loss-based methods inadequately capture strong correlations between output dimensions. CNFs are efficient in sampling and inference, they can be trained with a likelihood-based objective, and CNFs, being generative flows, do not suffer from mode collapse or training instabilities. We provide an effective method to train continuous CNFs for binary problems and in particular, we apply these CNFs to super-resolution and vessel segmentation tasks demonstrating competitive performance on standard benchmark datasets in terms of likelihood and conventional metrics. When the output y is high-dimensional this is a particularly challenging task, and the practitioner is left with many design choices. Do we factorize the conditional? If not, do we model correlations with, say, a conditional random field (Prince, 2012)? Do we use a unimodal distribution? How fat should the tails be? Do we use an explicit likelihood at all, or use implicit methods (Mohamed & Rezende, 2015) such as a GAN (Goodfellow et al., 2014)? Do we quantize the output?