In this paper, we present an end-to-end training framework for building state-of-the-art end-to-end speech recognition systems. Our training system utilizes a cluster of Central Processing Units(CPUs) and Graphics Processing Units (GPUs). The entire data reading, large scale data augmentation, neural network parameter updates are all performed "on-the-fly". We use vocal tract length perturbation [1] and an acoustic simulator [2] for data augmentation. The processed features and labels are sent to the GPU cluster. The Horovod allreduce approach is employed to train neural network parameters. We evaluated the effectiveness of our system on the standard Librispeech corpus [3] and the 10,000-hr anonymized Bixby English dataset. Our end-to-end speech recognition system built using this training infrastructure showed a 2.44 % WER on test-clean of the LibriSpeech test set after applying shallow fusion with a Transformer language model (LM). For the proprietary English Bixby open domain test set, we obtained a WER of 7.92 % using a Bidirectional Full Attention (BFA) end-to-end model after applying shallow fusion with an RNN-LM. When the monotonic chunckwise attention (MoCha) based approach is employed for streaming speech recognition, we obtained a WER of 9.95 % on the same Bixby open domain test set.

We build on the dynamical systems approach to deep learning, where deep residual networks are idealized as continuous-time dynamical systems. Although theoretical foundations have been developed on the optimization side through mean-field optimal control theory, the function approximation properties of such models remain largely unexplored, especially when the dynamical systems are controlled by functions of low complexity. In this paper, we establish some basic results on the approximation capabilities of deep learning models in the form of dynamical systems. In particular, we derive general sufficient conditions for universal approximation of functions in $L^p$ using flow maps of dynamical systems, and we also deduce some results on their approximation rates for specific cases. Overall, these results reveal that composition function approximation through flow maps present a new paradigm in approximation theory and contributes to building a useful mathematical framework to investigate deep learning.

To simultaneously capture syntax and global semantics from a text corpus, we propose a new larger-context recurrent neural network (RNN) based language model, which extracts recurrent hierarchical semantic structure via a dynamic deep topic model to guide natural language generation. Moving beyond a conventional RNN based language model that ignores long-range word dependencies and sentence order, the proposed model captures not only intra-sentence word dependencies, but also temporal transitions between sentences and inter-sentence topic dependences. For inference, we develop a hybrid of stochastic-gradient MCMC and recurrent autoencoding variational Bayes. Experimental results on a variety of real-world text corpora demonstrate that the proposed model not only outperforms state-of-the-art larger-context RNN-based language models, but also learns interpretable recurrent multilayer topics and generates diverse sentences and paragraphs that are syntactically correct and semantically coherent.

We introduce a novel autoencoder model that deviates from traditional autoen-coders by using the full latent vector to independently modulate each layer in the decoder. We demonstrate how such an'automodulator' allows for a principled approach to enforce latent space disentanglement, mixing of latent codes, and a straightforward way to utilize prior information that can be construed as a scale-specific invariance. Unlike the GAN models without encoders, autoencoder models can directly operate on new real input samples. This makes our model directly suitable for applications involving real-world inputs. As the architectural backbone, we extend recent generative autoencoder models that retain input identity and image sharpness at high resolutions better than V AEs. We show that our model achieves state-of-the-art latent space disentanglement and achieves high quality and diversity of output samples, as well as faithfulness of reconstructions. This paper introduces a new generative autoencoder for learning representations of image data sets, in a way that allows arbitrary combinations of latent codes to generate images (see Figure 1). We achieve this with an architecture that uses adaptive instance normalization (AdaIn, Dumoulin et al., 2017b; Huang & Belongie, 2017), and training methods that let the model learn a highly disentangled latent space by utilizing progressively growing autoencoders (Heljakka et al., 2019). In a typical autoencoder, input images are encoded into latent space, and the information of the latent variables is then passed through successive layers of decoding until a reconstruction of the input image has been formed. In our model, the latent vector independently modulates the statistics of each layer of the decoder--the output of layer n is no longer solely determined by the input from layer n 1. In image generation, the probability mass representing sensible images (such as human faces) lies concentrated on a low-dimensional manifold.

Heart failure hospitalization is a severe burden on healthcare. How to predict and therefore prevent readmission has been a significant challenge in outcomes research. To address this, we propose a deep learning approach to predict readmission from clinical notes. Unlike conventional methods that use structured data for prediction, we leverage the unstructured clinical notes to train deep learning models based on convolutional neural networks (CNN). We then use the trained models to classify and predict potentially high-risk admissions/patients. For evaluation, we trained CNNs using the discharge summary notes in the MIMIC III database. We also trained regular machine learning models based on random forest using the same datasets. The result shows that deep learning models outperform the regular models in prediction tasks. CNN method achieves a F1 score of 0.756 in general readmission prediction and 0.733 in 30-day readmission prediction, while random forest only achieves a F1 score of 0.674 and 0.656 respectively. We also propose a chi-square test based method to interpret key features associated with deep learning predicted readmissions. It reveals clinical insights about readmission embedded in the clinical notes. Collectively, our method can make the human evaluation process more efficient and potentially facilitate the reduction of readmission rates.

Deep convolutional neural networks are known to specialize in distilling compact and robust prior from a large amount of data. We are interested in applying deep networks in the absence of training dataset. In this paper, we introduce deep audio prior (DAP) which leverages the structure of a network and the temporal information in a single audio file. Specifically, we demonstrate that a randomly-initialized neural network can be used with carefully designed audio prior to tackle challenging audio problems such as universal blind source separation, interactive audio editing, audio texture synthesis, and audio co-separation. To understand the robustness of the deep audio prior, we construct a benchmark dataset \emph{Universal-150} for universal sound source separation with a diverse set of sources. We show superior audio results than previous work on both qualitative and quantitative evaluations. We also perform thorough ablation study to validate our design choices.

Deep neural networks (DNNs) perform well on a variety of tasks despite the fact that most networks used in practice are vastly overparametrized and even capable of perfectly fitting randomly labeled data. Recent evidence suggests that developing compressible representations is key for adjusting the complexity of overparametrized networks to the task at hand. In this paper, we provide new empirical evidence that supports this hypothesis by identifying two types of units that emerge when the network's width is increased: removable units which can be dropped out of the network without significant change to the output and repeated units whose activities are highly correlated with other units. The emergence of these units implies capacity constraints as the function the network represents could be expressed by a smaller network without these units. In a series of experiments with AlexNet, ResNet and Inception networks in the CIFAR-10 and ImageNet datasets, and also using shallow networks with synthetic data, we show that DNNs consistently increase either the number of removable units, repeated units, or both at greater widths for a comprehensive set of hyperparameters. These results suggest that the mechanisms by which networks in the deep learning regime adjust their complexity operate at the unit level and highlight the need for additional research into what drives the emergence of such units.

The field of machine learning has focused, primarily, on discretized sub-problems (i.e. vision, speech, natural language) of intelligence. While neuroscience tends to be observation heavy, providing few guiding theories. It is unlikely that artificial intelligence will emerge through only one of these disciplines. Instead, it is likely to be some amalgamation of their algorithmic and observational findings. As a result, there are a number of problems that should be addressed in order to select the beneficial aspects of both fields. In this article, we propose leading questions to guide the future of artificial intelligence research. There are clear computational principles on which the brain operates. The problem is finding these computational needles in a haystack of biological complexity. Biology has clear constraints but by not using it as a guide we are constraining ourselves.

The history of Convolutional neural networks have a remote origin. It is actually in 1979, when Professor Kunihiko Fukushima proposed a hierarchical, multilayered artificial neural network called The neocognitron. The neocognitron has been used for solving the problem of handwritten character recognition and some other pattern recognition tasks, and served as the inspiration for convolutional neural networks. But, if you asked about the history of the neocognitron, we simply can tell you that it was inspired by the model proposed by Hubel & Wiesel in 1959. They found two types of cells in the visual primary cortex called simple cells and complex cells, and also proposed a cascading model of these two types of cells for use in pattern recognition tasks.

Knowledge distillation is a model compression technique whereby a small network (student) is taught by a larger trained neural network (teacher). The smaller network is trained to behave like the large neural network. This enables the deployment of such models on small devices such as mobile phones or other edge devices. In this guide, we'll look at a couple of papers that attempt to tackle this challenge. In this paper, a small model is trained to generalize in the same way as the larger teacher model.