Optimizing Neural Networks via Koopman Operator Theory

Koopman operator theory, a powerful framework for discovering the underlying dynamics of nonlinear dynamical systems, was recently shown to be intimately connected with neural network training. In this work, we take the first steps in making use of this connection. As Koopman operator theory is a linear theory, a successful implementation of it in evolving network weights and biases offers the promise of accelerated training, especially in the context of deep networks, where optimization is inherently a non-convex problem. We show that Koopman operator theory methods allow for accurate predictions of the weights and biases of a feedforward, fully connected deep network over a non-trivial range of training time. During this time window, we find that our approach is at least 10x faster than gradient descent based methods, in line with the results expected from our complexity analysis. We highlight additional methods by which our results can be expanded to broader classes of networks and larger time intervals, which shall be the focus of future work in this novel intersection between dynamical systems and neural network theory.

RussianSuperGLUE: A Russian Language Understanding Evaluation Benchmark

In this paper, we introduce an advanced Russian general language understanding evaluation benchmark -- RussianGLUE. Recent advances in the field of universal language models and transformers require the development of a methodology for their broad diagnostics and testing for general intellectual skills - detection of natural language inference, commonsense reasoning, ability to perform simple logical operations regardless of text subject or lexicon. For the first time, a benchmark of nine tasks, collected and organized analogically to the SuperGLUE methodology, was developed from scratch for the Russian language. We provide baselines, human level evaluation, an open-source framework for evaluating models (https://github.com/RussianNLP/RussianSuperGLUE), and an overall leaderboard of transformer models for the Russian language. Besides, we present the first results of comparing multilingual models in the adapted diagnostic test set and offer the first steps to further expanding or assessing state-of-the-art models independently of language.

Textual Supervision for Visually Grounded Spoken Language Understanding

Visually-grounded models of spoken language understanding extract semantic information directly from speech, without relying on transcriptions. This is useful for low-resource languages, where transcriptions can be expensive or impossible to obtain. Recent work showed that these models can be improved if transcriptions are available at training time. However, it is not clear how an end-to-end approach compares to a traditional pipeline-based approach when one has access to transcriptions. Comparing different strategies, we find that the pipeline approach works better when enough text is available. With low-resource languages in mind, we also show that translations can be effectively used in place of transcriptions but more data is needed to obtain similar results.

ECG Signal Super-resolution by Considering Reconstruction and Cardiac Arrhythmias Classification Loss

With recent advances in deep learning algorithms, computer-assisted healthcare services have rapidly grown, especially for those that combine with mobile devices. Such a combination enables wearable and portable services for continuous measurements and facilitates real-time disease alarm based on physiological signals, e.g., cardiac arrhythmias (CAs) from electrocardiography (ECG). However, long-term and continuous monitoring confronts challenges arising from limitations of batteries, and the transmission bandwidth of devices. Therefore, identifying an effective way to improve ECG data transmission and storage efficiency has become an emerging topic. In this study, we proposed a deep-learning-based ECG signal super-resolution framework (termed ESRNet) to recover compressed ECG signals by considering the joint effect of signal reconstruction and CA classification accuracies. In our experiments, we downsampled the ECG signals from the CPSC 2018 dataset and subsequently evaluated the super-resolution performance by both reconstruction errors and classification accuracies. Experimental results showed that the proposed ESRNet framework can well reconstruct ECG signals from the 10-times compressed ones. Moreover, approximately half of the CA recognition accuracies were maintained within the ECG signals recovered by the ESRNet. The promising results confirm that the proposed ESRNet framework can be suitably used as a front-end process to reconstruct compressed ECG signals in real-world CA recognition scenarios.

Pose-Guided Human Animation from a Single Image in the Wild

We present a new pose transfer method for synthesizing a human animation from a single image of a person controlled by a sequence of body poses. Existing pose transfer methods exhibit significant visual artifacts when applying to a novel scene, resulting in temporal inconsistency and failures in preserving the identity and textures of the person. To address these limitations, we design a compositional neural network that predicts the silhouette, garment labels, and textures. Each modular network is explicitly dedicated to a subtask that can be learned from the synthetic data. At the inference time, we utilize the trained network to produce a unified representation of appearance and its labels in UV coordinates, which remains constant across poses. The unified representation provides an incomplete yet strong guidance to generating the appearance in response to the pose change. We use the trained network to complete the appearance and render it with the background. With these strategies, we are able to synthesize human animations that can preserve the identity and appearance of the person in a temporally coherent way without any fine-tuning of the network on the testing scene. Experiments show that our method outperforms the state-of-the-arts in terms of synthesis quality, temporal coherence, and generalization ability.

Do not let the history haunt you -- Mitigating Compounding Errors in Conversational Question Answering

The Conversational Question Answering (CoQA) task involves answering a sequence of inter-related conversational questions about a contextual paragraph. Although existing approaches employ human-written ground-truth answers for answering conversational questions at test time, in a realistic scenario, the CoQA model will not have any access to ground-truth answers for the previous questions, compelling the model to rely upon its own previously predicted answers for answering the subsequent questions. In this paper, we find that compounding errors occur when using previously predicted answers at test time, significantly lowering the performance of CoQA systems. To solve this problem, we propose a sampling strategy that dynamically selects between target answers and model predictions during training, thereby closely simulating the situation at test time. Further, we analyse the severity of this phenomena as a function of the question type, conversation length and domain type.

Optical Phase Dropout in Diffractive Deep Neural Network

Unitary learning is a backpropagation that serves to unitary weights update in deep complex-valued neural network with full connections, meeting a physical unitary prior in diffractive deep neural network ([DN]2). However, the square matrix property of unitary weights induces that the function signal has a limited dimension that could not generalize well. To address the overfitting problem that comes from the small samples loaded to [DN]2, an optical phase dropout trick is implemented. Phase dropout in unitary space that is evolved from a complex dropout and has a statistical inference is formulated for the first time. A synthetic mask recreated from random point apertures with random phase-shifting and its smothered modulation tailors the redundant links through incompletely sampling the input optical field at each diffractive layer. The physical features about the synthetic mask using different nonlinear activations are elucidated in detail. The equivalence between digital and diffractive model determines compound modulations that could successfully circumvent the nonlinear activations physically implemented in [DN]2. The numerical experiments verify the superiority of optical phase dropout in [DN]2 to enhance accuracy in 2D classification and recognition tasks-oriented.

The Macroeconomy as a Random Forest

Over the last decades, an impressive amount of non-linearities have been proposed to reconcile reduced-form macroeconomic models with the data. Many of them boil down to have linear regression coefficients evolving through time: threshold/switching/smooth-transition regression; structural breaks and random walk time-varying parameters. While all of these schemes are reasonably plausible in isolation, I argue that those are much more in agreement with the data if they are combined. To this end, I propose Macroeconomic Random Forests, which adapts the canonical Machine Learning (ML) algorithm to the problem of flexibly modeling evolving parameters in a linear macro equation. The approach exhibits clear forecasting gains over a wide range of alternatives and successfully predicts the drastic 2008 rise in unemployment. The obtained generalized time-varying parameters (GTVPs) are shown to behave differently compared to random walk coefficients by adapting nicely to the problem at hand, whether it is regime-switching behavior or long-run structural change. By dividing the typical ML interpretation burden into looking at each TVP separately, I find that the resulting forecasts are, in fact, quite interpretable. An application to the US Phillips curve reveals it is probably not flattening the way you think.

Addressing machine learning concept drift reveals declining vaccine sentiment during the COVID-19 pandemic

Social media analysis has become a common approach to assess public opinion on various topics, including those about health, in near real-time. The growing volume of social media posts has led to an increased usage of modern machine learning methods in natural language processing. While the rapid dynamics of social media can capture underlying trends quickly, it also poses a technical problem: algorithms trained on annotated data in the past may underperform when applied to contemporary data. This phenomenon, known as concept drift, can be particularly problematic when rapid shifts occur either in the topic of interest itself, or in the way the topic is discussed. Here, we explore the effect of machine learning concept drift by focussing on vaccine sentiments expressed on Twitter, a topic of central importance especially during the COVID-19 pandemic. We show that while vaccine sentiment has declined considerably during the COVID-19 pandemic in 2020, algorithms trained on pre-pandemic data would have largely missed this decline due to concept drift. Our results suggest that social media analysis systems must address concept drift in a continuous fashion in order to avoid the risk of systematic misclassification of data, which is particularly likely during a crisis when the underlying data can change suddenly and rapidly.

Online Exemplar Fine-Tuning for Image-to-Image Translation

Existing techniques to solve exemplar-based image-to-image translation within deep convolutional neural networks (CNNs) generally require a training phase to optimize the network parameters on domain-specific and task-specific benchmarks, thus having limited applicability and generalization ability. In this paper, we propose a novel framework, for the first time, to solve exemplar-based translation through an online optimization given an input image pair, called online exemplar fine-tuning (OEFT), in which we fine-tune the off-the-shelf and general-purpose networks to the input image pair themselves. We design two sub-networks, namely correspondence fine-tuning and multiple GAN inversion, and optimize these network parameters and latent codes, starting from the pre-trained ones, with well-defined loss functions. Our framework does not require the off-line training phase, which has been the main challenge of existing methods, but the pre-trained networks to enable optimization in online. Experimental results prove that our framework is effective in having a generalization power to unseen image pairs and clearly even outperforms the state-of-the-arts needing the intensive training phase.