This paper presents FastFit, a novel neural vocoder architecture that replaces the U-Net encoder with multiple short-time Fourier transforms (STFTs) to achieve faster generation rates without sacrificing sample quality. We replaced each encoder block with an STFT, with parameters equal to the temporal resolution of each decoder block, leading to the skip connection. FastFit reduces the number of parameters and the generation time of the model by almost half while maintaining high fidelity. Through objective and subjective evaluations, we demonstrated that the proposed model achieves nearly twice the generation speed of baseline iteration-based vocoders while maintaining high sound quality. We further showed that FastFit produces sound qualities similar to those of other baselines in text-to-speech evaluation scenarios, including multi-speaker and zero-shot text-to-speech.
Applications of normalizing flows to the sampling of field configurations in lattice gauge theory have so far been explored almost exclusively in two space-time dimensions. We report new algorithmic developments of gauge-equivariant flow architectures facilitating the generalization to higher-dimensional lattice geometries. Specifically, we discuss masked autoregressive transformations with tractable and unbiased Jacobian determinants, a key ingredient for scalable and asymptotically exact flow-based sampling algorithms. For concreteness, results from a proof-of-principle application to SU(3) lattice gauge theory in four space-time dimensions are reported.
Despite the broad application of Machine Learning models as a Service (MLaaS), they are vulnerable to model stealing attacks. These attacks can replicate the model functionality by using the black-box query process without any prior knowledge of the target victim model. Existing stealing defenses add deceptive perturbations to the victim's posterior probabilities to mislead the attackers. However, these defenses are now suffering problems of high inference computational overheads and unfavorable trade-offs between benign accuracy and stealing robustness, which challenges the feasibility of deployed models in practice. To address the problems, this paper proposes Isolation and Induction (InI), a novel and effective training framework for model stealing defenses. Instead of deploying auxiliary defense modules that introduce redundant inference time, InI directly trains a defensive model by isolating the adversary's training gradient from the expected gradient, which can effectively reduce the inference computational cost. In contrast to adding perturbations over model predictions that harm the benign accuracy, we train models to produce uninformative outputs against stealing queries, which can induce the adversary to extract little useful knowledge from victim models with minimal impact on the benign performance. Extensive experiments on several visual classification datasets (e.g., MNIST and CIFAR10) demonstrate the superior robustness (up to 48% reduction on stealing accuracy) and speed (up to 25.4x faster) of our InI over other state-of-the-art methods. Our codes can be found in https://github.com/DIG-Beihang/InI-Model-Stealing-Defense.
DNN accelerators have been widely deployed in many scenarios to speed up the inference process and reduce the energy consumption. One big concern about the usage of the accelerators is the confidentiality of the deployed models: model inference execution on the accelerators could leak side-channel information, which enables an adversary to preciously recover the model details. Such model extraction attacks can not only compromise the intellectual property of DNN models, but also facilitate some adversarial attacks. Although previous works have demonstrated a number of side-channel techniques to extract models from DNN accelerators, they are not practical for two reasons. (1) They only target simplified accelerator implementations, which have limited practicality in the real world. (2) They require heavy human analysis and domain knowledge. To overcome these limitations, this paper presents Mercury, the first automated remote side-channel attack against the off-the-shelf Nvidia DNN accelerator. The key insight of Mercury is to model the side-channel extraction process as a sequence-to-sequence problem. The adversary can leverage a time-to-digital converter (TDC) to remotely collect the power trace of the target model's inference. Then he uses a learning model to automatically recover the architecture details of the victim model from the power trace without any prior knowledge. The adversary can further use the attention mechanism to localize the leakage points that contribute most to the attack. Evaluation results indicate that Mercury can keep the error rate of model extraction below 1%.
Real-time bidding has emerged as an effective online advertising technique. With real-time bidding, advertisers can position ads per impression, enabling them to optimise ad campaigns by targeting specific audiences in real-time. This paper proposes a novel method for real-time bidding that combines deep learning and reinforcement learning techniques to enhance the efficiency and precision of the bidding process. In particular, the proposed method employs a deep neural network to predict auction details and market prices and a reinforcement learning algorithm to determine the optimal bid price. The model is trained using historical data from the iPinYou dataset and compared to cutting-edge real-time bidding algorithms. The outcomes demonstrate that the proposed method is preferable regarding cost-effectiveness and precision. In addition, the study investigates the influence of various model parameters on the performance of the proposed algorithm. It offers insights into the efficacy of the combined deep learning and reinforcement learning approach for real-time bidding. This study contributes to advancing techniques and offers a promising direction for future research.
Open-vocabulary segmentation is a challenging task requiring segmenting and recognizing objects from an open set of categories. One way to address this challenge is to leverage multi-modal models, such as CLIP, to provide image and text features in a shared embedding space, which bridges the gap between closed-vocabulary and open-vocabulary recognition. Hence, existing methods often adopt a two-stage framework to tackle the problem, where the inputs first go through a mask generator and then through the CLIP model along with the predicted masks. This process involves extracting features from images multiple times, which can be ineffective and inefficient. By contrast, we propose to build everything into a single-stage framework using a shared Frozen Convolutional CLIP backbone, which not only significantly simplifies the current two-stage pipeline, but also remarkably yields a better accuracy-cost trade-off. The proposed FC-CLIP, benefits from the following observations: the frozen CLIP backbone maintains the ability of open-vocabulary classification and can also serve as a strong mask generator, and the convolutional CLIP generalizes well to a larger input resolution than the one used during contrastive image-text pretraining. When training on COCO panoptic data only and testing in a zero-shot manner, FC-CLIP achieve 26.8 PQ, 16.8 AP, and 34.1 mIoU on ADE20K, 18.2 PQ, 27.9 mIoU on Mapillary Vistas, 44.0 PQ, 26.8 AP, 56.2 mIoU on Cityscapes, outperforming the prior art by +4.2 PQ, +2.4 AP, +4.2 mIoU on ADE20K, +4.0 PQ on Mapillary Vistas and +20.1 PQ on Cityscapes, respectively. Additionally, the training and testing time of FC-CLIP is 7.5x and 6.6x significantly faster than the same prior art, while using 5.9x fewer parameters. FC-CLIP also sets a new state-of-the-art performance across various open-vocabulary semantic segmentation datasets. Code at https://github.com/bytedance/fc-clip
Stress is widely recognized as a major contributor to a variety of health issues. Stress prediction using biosignal data recorded by wearables is a key area of study in mobile sensing research because real-time stress prediction can enable digital interventions to immediately react at the onset of stress, helping to avoid many psychological and physiological symptoms such as heart rhythm irregularities. Electrodermal activity (EDA) is often used to measure stress. However, major challenges with the prediction of stress using machine learning include the subjectivity and sparseness of the labels, a large feature space, relatively few labels, and a complex nonlinear and subjective relationship between the features and outcomes. To tackle these issues, we examine the use of model personalization: training a separate stress prediction model for each user. To allow the neural network to learn the temporal dynamics of each individual's baseline biosignal patterns, thus enabling personalization with very few labels, we pre-train a 1-dimensional convolutional neural network (CNN) using self-supervised learning (SSL). We evaluate our method using the Wearable Stress and Affect prediction (WESAD) dataset. We fine-tune the pre-trained networks to the stress prediction task and compare against equivalent models without any self-supervised pre-training. We discover that embeddings learned using our pre-training method outperform supervised baselines with significantly fewer labeled data points: the models trained with SSL require less than 30% of the labels to reach equivalent performance without personalized SSL. This personalized learning method can enable precision health systems which are tailored to each subject and require few annotations by the end user, thus allowing for the mobile sensing of increasingly complex, heterogeneous, and subjective outcomes such as stress.
Low-lying coastal cities, exemplified by Norfolk, Virginia, face the challenge of street flooding caused by rainfall and tides, which strain transportation and sewer systems and can lead to property damage. While high-fidelity, physics-based simulations provide accurate predictions of urban pluvial flooding, their computational complexity renders them unsuitable for real-time applications. Using data from Norfolk rainfall events between 2016 and 2018, this study compares the performance of a previous surrogate model based on a random forest algorithm with two deep learning models: Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). This investigation underscores the importance of using a model architecture that supports the communication of prediction uncertainty and the effective integration of relevant, multi-modal features.
The optimization of network performance is vital for the delivery of services using standard cellular technologies for mobile communications. Call setup delay and User Equipment (UE) battery savings significantly influence network performance. Improving these factors is vital for ensuring optimal service delivery. In comparison to traditional circuit-switched voice calls, VoLTE (Voice over LTE) technology offers faster call setup durations and better battery-saving performance. To validate these claims, a drive test was carried out using the XCAL drive test tool to collect real-time network parameter details in VoLTE and non-VoLTE voice calls. The findings highlight the analysis of real-time network characteristics, such as the call setup delay calculation, battery-saving performance, and DRX mechanism. The study contributes to the understanding of network optimization strategies and provides insights for enhancing the quality of service (QoS) in mobile communication networks. Examining VoLTE and non-VoLTE operations, this research highlights the substantial energy savings obtained by VoLTE. Specifically, VoLTE saves approximately 60.76% of energy before the Service Request and approximately 38.97% of energy after the Service Request. Moreover, VoLTE to VoLTE calls have a 72.6% faster call setup delay than non-VoLTE-based LTE to LTE calls, because of fewer signaling messages required. Furthermore, as compared to non-VoLTE to non-VoLTE calls, VoLTE to non-VoLTE calls offer an 18.6% faster call setup delay. These results showcase the performance advantages of VoLTE and reinforce its potential for offering better services in wireless communication networks.
We consider the supervised learning problem of learning the price of an option or the implied volatility given appropriate input data (model parameters) and corresponding output data (option prices or implied volatilities). The majority of articles in this literature considers a (plain) feed forward neural network architecture in order to connect the neurons used for learning the function mapping inputs to outputs. In this article, motivated by methods in image classification and recent advances in machine learning methods for PDEs, we investigate empirically whether and how the choice of network architecture affects the accuracy and training time of a machine learning algorithm. We find that for option pricing problems, where we focus on the Black--Scholes and the Heston model, the generalized highway network architecture outperforms all other variants, when considering the mean squared error and the training time as criteria. Moreover, for the computation of the implied volatility, after a necessary transformation, a variant of the DGM architecture outperforms all other variants, when considering again the mean squared error and the training time as criteria.