Articulatory distinctive features, as well as phonetic transcription, play important role in speech-related tasks: computer-assisted pronunciation training, text-to-speech conversion (TTS), studying speech production mechanisms, speech recognition for low-resourced languages. End-to-end approaches to speech-related tasks got a lot of traction in recent years. We apply Listen, Attend and Spell~(LAS)~\cite{Chan-LAS2016} architecture to phones recognition on a small small training set, like TIMIT~\cite{TIMIT-1992}. Also, we introduce a novel decoding technique that allows to train manners and places of articulation detectors end-to-end using attention models. We also explore joint phones recognition and articulatory features detection in multitask learning setting.

Conventional sequential learning methods such as Recurrent Neural Networks (RNNs) focus on interactions between consecutive inputs, i.e. first-order Markovian dependency. However, most of sequential data, as seen with videos, have complex temporal dependencies that imply variable-length semantic flows and their compositions, and those are hard to be captured by conventional methods. Here, we propose Temporal Dependency Networks (TDNs) for learning video data by discovering these complex structures of the videos. The TDNs represent video as a graph whose nodes and edges correspond to frames of the video and their dependencies respectively. Via a parameterized kernel with graph-cut and graph convolutions, the TDNs find compositional temporal dependencies of the data in multilevel graph forms. We evaluate the proposed method on the large-scale video dataset Youtube-8M. The experimental results show that our model efficiently learns the complex semantic structure of video data.

Graph embedding is a popular algorithmic approach for creating vector representations for individual vertices in networks. Training these algorithms at scale is important for creating embeddings that can be used for classification, ranking, recommendation and other common applications in industry. While industrial systems exist for training graph embeddings on large datasets, many of these distributed architectures are forced to partition copious amounts of data and model logic across many worker nodes. In this paper, we propose a distributed infrastructure that completely avoids graph partitioning, dynamically creates size constrained computational graphs across worker nodes, and uses highly efficient indexing operations for updating embeddings that allow the system to function at scale. We show that our system can scale an existing embeddings algorithm - skip-gram - to train on the open-source Friendster network (68 million vertices) and on an internal heterogeneous graph (50 million vertices). We measure the performance of our system on two key quantitative metrics: link-prediction accuracy and rate of convergence. We conclude this work by analyzing how a greater number of worker nodes actually improves our system's performance on the aforementioned metrics and discuss our next steps for rigorously evaluating the embedding vectors produced by our system.

Anomaly detection is a very worthwhile question. However, the anomaly is not a simple two-category in reality, so it is difficult to give accurate results through the comparison of similarities. There are already some deep learning models based on GAN for anomaly detection that demonstrate validity and accuracy on time series data sets. In this paper, we propose an unsupervised model-based anomaly detection named LVEAD, which assumpts that the anomalies are objects that do not fit perfectly with the model. For better handling the time series, we use the LSTM model as the encoder and decoder part of the VAE model. Considering to better distinguish the normal and anomaly data, we train a re-encoder model to the latent space to generate new data. Experimental results of several benchmarks show that our method outperforms state-of-the-art anomaly detection techniques.

The performance of deep neural networks is highly sensitive to the choice of the hyperparameters that define the structure of the network and the learning process. When facing a new application, tuning a deep neural network is a tedious and time consuming process that is often described as a "dark art". This explains the necessity of automating the calibration of these hyperparameters. Derivative-free optimization is a field that develops methods designed to optimize time consuming functions without relying on derivatives. This work introduces the HyperNOMAD package, an extension of the NOMAD software that applies the MADS algorithm [7] to simultaneously tune the hyperparameters responsible for both the architecture and the learning process of a deep neural network (DNN), and that allows for an important flexibility in the exploration of the search space by taking advantage of categorical variables. This new approach is tested on the MNIST and CIFAR-10 data sets and achieves results comparable to the current state of the art.

We design a new robust clustering algorithm that can deal efficiently with noise and outliers in diverse datasets. As an EM-like algorithm, it is based on both estimations of clusters centers and covariances but also on a scale parameter per data-point. This allows the algorithm to accommodate for heavier/lighter tails distributions (in comparison to classical Gaussian distributions) and outliers without significantly loosing efficiency in classical scenarios. Convergence and accuracy of the algorithm are first analyzed by considering synthetic data. Then, we show that the proposed algorithm outperforms other classical unsupervised methods of the literature such as k-means, the EM algorithm and HDBSCAN when applied to real datasets as MNIST, NORB and 20newsgroups.

Many manifold embedding algorithms fail apparently when the data manifold has a large aspect ratio (such as a long, thin strip). Here, we formulate success and failure in terms of finding a smooth embedding, showing also that the problem is pervasive and more complex than previously recognized. Mathematically, success is possible under very broad conditions, provided that embedding is done by carefully selected eigenfunctions of the Laplace-Beltrami operator $\Delta$. Hence, we propose a bicriterial Independent Eigencoordinate Selection (IES) algorithm that selects smooth embeddings with few eigenvectors. The algorithm is grounded in theory, has low computational overhead, and is successful on synthetic and large real data.

Deep Neural Networks are robust to minor perturbations of the learned network parameters and their minor modifications do not change the overall network response significantly. This allows space for model stealing, where a malevolent attacker can steal an already trained network, modify the weights and claim the new network his own intellectual property. In certain cases this can prevent the free distribution and application of networks in the embedded domain. In this paper, we propose a method for creating an equivalent version of an already trained fully connected deep neural network that can prevent network stealing: namely, it produces the same responses and classification accuracy, but it is extremely sensitive to weight changes.

An $\varepsilon$-coreset for Least-Mean-Squares (LMS) of a matrix $A\in{\mathbb{R}}^{n\times d}$ is a small weighted subset of its rows that approximates the sum of squared distances from its rows to every affine $k$-dimensional subspace of ${\mathbb{R}}^d$, up to a factor of $1\pm\varepsilon$. Such coresets are useful for hyper-parameter tuning and solving many least-mean-squares problems such as low-rank approximation ($k$-SVD), $k$-PCA, Lassso/Ridge/Linear regression and many more. Coresets are also useful for handling streaming, dynamic and distributed big data in parallel. With high probability, non-uniform sampling based on upper bounds on what is known as importance or sensitivity of each row in $A$ yields a coreset. The size of the (sampled) coreset is then near-linear in the total sum of these sensitivity bounds. We provide algorithms that compute provably \emph{tight} bounds for the sensitivity of each input row. It is based on two ingredients: (i) iterative algorithm that computes the exact sensitivity of each point up to arbitrary small precision for (non-affine) $k$-subspaces, and (ii) a general reduction of independent interest from computing sensitivity for the family of affine $k$-subspaces in ${\mathbb{R}}^d$ to (non-affine) $(k+1)$- subspaces in ${\mathbb{R}}^{d+1}$. Experimental results on real-world datasets, including the English Wikipedia documents-term matrix, show that our bounds provide significantly smaller and data-dependent coresets also in practice. Full open source is also provided.

Transfer learning focuses on the reuse of supervised learning models in a new context. Prominent applications can be found in robotics, image processing or web mining. In these areas, learning scenarios change by nature, but often remain related and motivate the reuse of existing supervised models. While the majority of symmetric and asymmetric domain adaptation algorithms utilize all available source and target domain data, we show that domain adaptation requires only a substantial smaller subset. This makes it more suitable for real-world scenarios where target domain data is rare. The presented approach finds a target subspace representation for source and target data to address domain differences by orthogonal basis transfer. We employ Nystr\"om techniques and show the reliability of this approximation without a particular landmark matrix by applying post-transfer normalization. It is evaluated on typical domain adaptation tasks with standard benchmark data.