The verification problem for neural networks is verifying whether a neural network will suffer from adversarial samples, or approximating the maximal allowed scale of adversarial perturbation that can be endured. While most prior work contributes to verifying feed-forward networks, little has been explored for verifying recurrent networks. This is due to the existence of a more rigorous constraint on the perturbation space for sequential data, and the lack of a proper metric for measuring the perturbation. In this work, we address these challenges by proposing a metric which measures the distance between strings, and use deterministic finite automata (DFA) to represent a rigorous oracle which examines if the generated adversarial samples violate certain constraints on a perturbation. More specifically, we empirically show that certain recurrent networks allow relatively stable DFA extraction. As such, DFAs extracted from these recurrent networks can serve as a surrogate oracle for when the ground truth DFA is unknown. We apply our verification mechanism to several widely used recurrent networks on a set of the Tomita grammars. The results demonstrate that only a few models remain robust against adversarial samples. In addition, we show that for grammars with different levels of complexity, there is also a difference in the difficulty of robust learning of these grammars.

Due to the increasing complexity of Integrated Circuits (ICs) and System-on-Chip (SoC), developing high-quality synthesis flows within a short market time becomes more challenging. We propose a general approach that precisely estimates the Quality-of-Result (QoR), such as delay and area, of unseen synthesis flows for specific designs. The main idea is training a Recurrent Neural Network (RNN) regressor, where the flows are inputs and QoRs are ground truth. The RNN regressor is constructed with Long Short-Term Memory (LSTM) and fully-connected layers. This approach is demonstrated with 1.2 million data points collected using 14nm, 7nm regular-voltage (RVT), and 7nm low-voltage (LVT) FinFET technologies with twelve IC designs. The accuracy of predicting the QoRs (delay and area) within one technology is $\boldsymbol{\geq}$\textbf{98.0}\% over $\sim$240,000 test points. To enable accurate predictions cross different technologies and different IC designs, we propose a transfer-learning approach that utilizes the model pre-trained with 14nm datasets. Our transfer learning approach obtains estimation accuracy $\geq$96.3\% over $\sim$960,000 test points, using only 100 data points for training.

One of the most important problems of data processing in high energy and nuclear physics is the event reconstruction. Its main part is the track reconstruction procedure which consists in looking for all tracks that elementary particles leave when they pass through a detector among a huge number of points, so-called hits, produced when flying particles fire detector coordinate planes. Unfortunately, the tracking is seriously impeded by the famous shortcoming of multiwired, strip and GEM detectors due to appearance in them a lot of fake hits caused by extra spurious crossings of fired strips. Since the number of those fakes is several orders of magnitude greater than for true hits, one faces with the quite serious difficulty to unravel possible track-candidates via true hits ignoring fakes. We introduce a renewed method that is a significant improvement of our previous two-stage approach based on hit preprocessing using directed K-d tree search followed a deep neural classifier. We combine these two stages in one by applying recurrent neural network that simultaneously determines whether a set of points belongs to a true track or not and predicts where to look for the next point of track on the next coordinate plane of the detector. We show that proposed deep network is more accurate, faster and does not require any special preprocessing stage. Preliminary results of our approach for simulated events of the BM@N GEM detector are presented.

Typical large-scale recommender systems use deep learning models that are stored on a large amount of DRAM. These models often rely on embeddings, which consume most of the required memory. We present Bandana, a storage system that reduces the DRAM footprint of embeddings, by using Nonvolatile Memory (NVM) as the primary storage medium, with a small amount of DRAM as cache. The main challenge in storing embeddings on NVM is its limited read bandwidth compared to DRAM. Bandana uses two primary techniques to address this limitation: first, it stores embedding vectors that are likely to be read together in the same physical location, using hypergraph partitioning, and second, it decides the number of embedding vectors to cache in DRAM by simulating dozens of small caches. These techniques allow Bandana to increase the effective read bandwidth of NVM by 2-3 and thereby significantly reduce the total cost of ownership. An increasing number of web-scale applications are relying on deep learning models, including online search (Clark, 2015), online ads (Zhou et al., 2018) and content recommendation systems (Covington et al., 2016). Therefore, application providers are dedicating ever more compute and storage resources to training, storing and accessing deep learning models. For example, at Facebook, thousands of servers are dedicated to storing deep learning models to recommend relevant posts or content to users. The deep learning features are often represented by vectors called embeddings, which encode the meaning of each feature, such that similar vectors are closer in the embedding Euclidean space. To compute the most relevant posts to serve each user, Facebook uses two types of embeddings: user and post embeddings.

Characterizing statistical properties of solutions of inverse problems is essential for decision making. Bayesian inversion offers a tractable framework for this purpose, but current approaches are computationally unfeasible for most realistic imaging applications in the clinic. We introduce two novel deep learning based methods for solving large-scale inverse problems using Bayesian inversion: a sampling based method using a WGAN with a novel mini-discriminator and a direct approach that trains a neural network using a novel loss function. The performance of both methods is demonstrated on image reconstruction in ultra low dose 3D helical CT. We compute the posterior mean and standard deviation of the 3D images followed by a hypothesis test to assess whether a "dark spot" in the liver of a cancer stricken patient is present. Both methods are computationally efficient and our evaluation shows very promising performance that clearly supports the claim that Bayesian inversion is usable for 3D imaging in time critical applications.

In this work, sequence-to-sequence (seq2seq) models, originally developed for language translation, are used to predict the temporal evolution of complex, multi-physics computer simulations. The predictive performance of seq2seq models is compared to state transition models for datasets generated with multiphysics codes with varying levels of complexity-from simple 1D diffusion calculations to simulations of inertial confinement fusion implosions. Seq2seq models demonstrate the ability to accurately emulate complex systems, enabling the rapid estimation of the evolution of quantities of interest in computationally expensive simulations. Keywords: recurrent neural network, sequence-to-sequence, multiphysics simulation, radiation hydrodynamics 1. Introduction Computer simulations of detailed multi-physics systems often take several hours to run, making exploration throughout a vast design space prohibitively expensive. A common method for mapping design spaces of large computer codes is to train a machine learning model to emulate the code in a region of the design space [1, 2, 3]. The machine learning model, often called a "surrogate", D. Humbird) Preprint submitted to Elsevier November 15, 2018 learns to accurately interpolate between a set of simulations spread throughout the reduced design space, such that it can rapidly predict quantities of interest anywhere within that space without requiring additional expensive simulations.

Overfitting frequently occurs in deep learning. In this paper, we propose a novel regularization method called Drop-Activation to reduce overfitting and improve generalization. The key idea is to \emph{drop} nonlinear activation functions by setting them to be identity functions randomly during training time. During testing, we use a deterministic network with a new activation function to encode the average effect of dropping activations randomly. Experimental results on CIFAR-10, CIFAR-100, SVHN, and EMNIST show that Drop-Activation generally improves the performance of popular neural network architectures. Furthermore, unlike dropout, as a regularizer Drop-Activation can be used in harmony with standard training and regularization techniques such as Batch Normalization and AutoAug. Our theoretical analyses support the regularization effect of Drop-Activation as implicit parameter reduction and its capability to be used together with Batch Normalization.

We propose a learning-based approach for estimating the spectrum of a multisinusoidal signal from a finite number of samples. A neural-network is trained to approximate the spectra of such signals on simulated data. The proposed methodology is very flexible: adapting to different signal and noise models only requires modifying the training data accordingly. Numerical experiments show that the approach performs competitively with classical methods designed for additive Gaussian noise at a range of noise levels, and is also effective in the presence of impulsive noise.

Developing accurate, transferable and computationally inexpensive machine learning models can rapidly accelerate the discovery and development of new materials. Some of the major challenges involved in developing such models are, (i) limited availability of materials data as compared to other fields, (ii) lack of universal descriptor of materials to predict its various properties. The limited availability of materials data can be addressed through transfer learning, while the generic representation was recently addressed by Xie and Grossman [1], where they developed a crystal graph convolutional neural network (CGCNN) that provides a unified representation of crystals. In this work, we develop a new model (MT-CGCNN) by integrating CGCNN with transfer learning based on multi-task (MT) learning. We demonstrate the effectiveness of MT-CGCNN by simultaneous prediction of various material properties such as Formation Energy, Band Gap and Fermi Energy for a wide range of inorganic crystals (46774 materials). MT-CGCNN is able to reduce the test error when employed on correlated properties by upto 8%. The model prediction has lower test error compared to CGCNN, even when the training data is reduced by 10%. We also demonstrate our model's better performance through prediction of end user scenario related to metal/non-metal classification. These results encourage further development of machine learning approaches which leverage multi-task learning to address the aforementioned challenges in the discovery of new materials. We make MT-CGCNN's source code available to encourage reproducible research.

Deep neural networks (DNNs) have demonstrated dominating performance in many fields; since AlexNet, the neural networks used in practice are going wider and deeper. On the theoretical side, a long line of works have been focusing on why we can train neural networks when there is only one hidden layer. The theory of multi-layer networks remains somewhat unsettled. In this work, we prove why simple algorithms such as stochastic gradient descent (SGD) can find $\textit{global minima}$ on the training objective of DNNs. We only make two assumptions: the inputs do not degenerate and the network is over-parameterized. The latter means the number of hidden neurons is sufficiently large: $\textit{polynomial}$ in $L$, the number of DNN layers and in $n$, the number of training samples. As concrete examples, on the training set and starting from randomly initialized weights, we show that SGD attains 100% accuracy in classification tasks, or minimizes regression loss in linear convergence speed $\varepsilon \propto e^{-\Omega(T)}$, with a number of iterations that only scales polynomial in $n$ and $L$. Our theory applies to the widely-used but non-smooth ReLU activation, and to any smooth and possibly non-convex loss functions. In terms of network architectures, our theory at least applies to fully-connected neural networks, convolutional neural networks (CNN), and residual neural networks (ResNet).