Amazon Web Services has included Amazon machine images or AMIs, which are master images for the creation of virtual servers, for deep learning in the AWS GovCloud, a cloud region built for sensitive data and regulated workloads. The AMIs are designed to provide a platform wherein developers would construct artificial intelligence applications based on deep learning frameworks such as TensorFlow, Apache MXNet and Gluon, PyTorch and Chainer, AWS said Thursday. The images can also be operated with Ubuntu and Linux platforms for clean slate or custom setups, and include NVIDIA GPU acceleration to help speed up model training. The AMIs do not require additional charges.

Convolutional neural networks (CNN) have been shown to provide a good solution for classification problems that utilize data obtained from vibrational spectroscopy. Moreover, CNNs are capable of identification from noisy spectra without the need for additional preprocessing. However, their application in practical spectroscopy is limited due to two shortcomings. The effectiveness of the classification using CNNs drops rapidly when only a small number of spectra per substance are available for training (which is a typical situation in real applications). Additionally, to accommodate new, previously unseen substance classes, the network must be retrained which is computationally intensive. Here we address these issues by reformulating a multi-class classification problem with a large number of classes, but a small number of samples per class, to a binary classification problem with sufficient data available for representation learning. Namely, we define the learning task as identifying pairs of inputs as belonging to the same or different classes. We achieve this using a Siamese convolutional neural network. A novel sampling strategy is proposed to address the imbalance problem in training the Siamese Network. The trained network can effectively classify samples of unseen substance classes using just a single reference sample (termed as one-shot learning in the machine learning community). Our results demonstrate better accuracy than other practical systems to date, while allowing effortless updates of the system's database with novel substance classes.

This paper addresses the scalability challenge of architecture search by formulating the task in a differentiable manner. Unlike conventional approaches of applying evolution or reinforcement learning over a discrete and non-differentiable search space, our method is based on the continuous relaxation of the architecture representation, allowing efficient search of the architecture using gradient descent.

Distributed model training suffers from communication overheads due to frequent gradient updates transmitted between compute nodes. To mitigate these overheads, several studies propose the use of sparsified stochastic gradients. We argue that these are facets of a general sparsification method that can operate on any possible atomic decomposition. Notable examples include element-wise, singular value, and Fourier decompositions. We present ATOMO, a general framework for atomic sparsification of stochastic gradients. Given a gradient, an atomic decomposition, and a sparsity budget, ATOMO gives a random unbiased sparsification of the atoms minimizing variance. We show that methods such as QSGD and TernGrad are special cases of ATOMO and show that sparsifiying gradients in their singular value decomposition (SVD), rather than the coordinate-wise one, can lead to significantly faster distributed training.

Recent years have witnessed the unprecedented rising of time series from almost all kindes of academic and industrial fields. Various types of deep neural network models have been introduced to time series analysis, but the important frequency information is yet lack of effective modeling. In light of this, in this paper we propose a wavelet-based neural network structure called multilevel Wavelet Decomposition Network (mWDN) for building frequency-aware deep learning models for time series analysis. mWDN preserves the advantage of multilevel discrete wavelet decomposition in frequency learning while enables the fine-tuning of all parameters under a deep neural network framework. Based on mWDN, we further propose two deep learning models called Residual Classification Flow (RCF) and multi-frequecy Long Short-Term Memory (mLSTM) for time series classification and forecasting, respectively. The two models take all or partial mWDN decomposed sub-series in different frequencies as input, and resort to the back propagation algorithm to learn all the parameters globally, which enables seamless embedding of wavelet-based frequency analysis into deep learning frameworks. Extensive experiments on 40 UCR datasets and a real-world user volume dataset demonstrate the excellent performance of our time series models based on mWDN. In particular, we propose an importance analysis method to mWDN based models, which successfully identifies those time-series elements and mWDN layers that are crucially important to time series analysis. This indeed indicates the interpretability advantage of mWDN, and can be viewed as an indepth exploration to interpretable deep learning.

The stroke sequence of characters is significant for the character recognition task. In this paper, we propose a stroke-based character recognition (SCR) method. We train a stroke inference module under deep reinforcement learning (DRL) framework. This module extracts the sequence of strokes from characters, which can be integrated with character recognizers to improve their robustness to noise. Our experiments show that the module can handle complicated noise and reconstruct the characters. Meanwhile, it can also help achieve great ability in defending adversarial attacks of character recognizers.

We propose a novel online algorithm for training deep feedforward neural networks that employs alternating minimization (block-coordinate descent) between the weights and activation variables. It extends off-line alternating minimization approaches to online, continual learning, and improves over stochastic gradient descent (SGD) with backpropagation in several ways: it avoids the vanishing gradient issue, it allows for non-differentiable nonlinearities, and it permits parallel weight updates across the layers. Unlike SGD, our approach employs co-activation memory inspired by the online sparse coding algorithm of [Mairal et al, 2009]. Furthermore, local iterative optimization with explicit activation updates is a potentially more biologically plausible learning mechanism than backpropagation. We provide theoretical convergence analysis and promising empirical results on several datasets.

The current paper is a study in Recurrent Neural Networks (RNN), motivated by the lack of examples simple enough so that they can be thoroughly understood theoretically, but complex enough to be realistic. We constructed an example of structured data, motivated by problems from image-to-text conversion (OCR), which requires long-term memory to decode. Our data is a simple writing system, encoding characters 'X' and 'O' as their upper halves, which is possible due to symmetry of the two characters. The characters can be connected, as in some languages using cursive, such as Arabic (abjad). The string 'XOOXXO' may be encoded as '${\vee}{\wedge}\kern-1.5pt{\wedge}{\vee}\kern-1.5pt{\vee}{\wedge}$'. It follows that we may need to know arbitrarily long past to decode a current character, thus requiring long-term memory. Subsequently we constructed an RNN capable of decoding sequences encoded in this manner. Rather than by training, we constructed our RNN "by inspection", i.e. we guessed its weights. This involved a sequence of steps. We wrote a conventional program which decodes the sequences as the example above. Subsequently, we interpreted the program as a neural network (the only example of this kind known to us). Finally, we generalized this neural network to discover a new RNN architecture whose instance is our handcrafted RNN. It turns out to be a 3 layer network, where the middle layer is capable of performing simple logical inferences; thus the name "deductron". It is demonstrated that it is possible to train our network by simulated annealing. Also, known variants of stochastic gradient descent (SGD) methods are shown to work.

Evaluating on adversarial examples has become a standard procedure to measure robustness of deep learning models. Due to the difficulty of creating white-box adversarial examples for discrete text input, most analyses of the robustness of NLP models have been done through black-box adversarial examples. We investigate adversarial examples for character-level neural machine translation (NMT), and contrast black-box adversaries with a novel white-box adversary, which employs differentiable string-edit operations to rank adversarial changes. We propose two novel types of attacks which aim to remove or change a word in a translation, rather than simply break the NMT. We demonstrate that white-box adversarial examples are significantly stronger than their black-box counterparts in different attack scenarios, which show more serious vulnerabilities than previously known. In addition, after performing adversarial training, which takes only 3 times longer than regular training, we can improve the model's robustness significantly.

To be informative, an evaluation must measure how well systems generalize to realistic unseen data. We identify limitations of and propose improvements to current evaluations of text-to-SQL systems. First, we compare human-generated and automatically generated questions, characterizing properties of queries necessary for real-world applications. To facilitate evaluation on multiple datasets, we release standardized and improved versions of seven existing datasets and one new text-to-SQL dataset. Second, we show that the current division of data into training and test sets measures robustness to variations in the way questions are asked, but only partially tests how well systems generalize to new queries; therefore, we propose a complementary dataset split for evaluation of future work. Finally, we demonstrate how the common practice of anonymizing variables during evaluation removes an important challenge of the task. Our observations highlight key difficulties, and our methodology enables effective measurement of future development.