Zhenjie Zhang

Zining Zhang, Bingsheng He, Zhenjie Zhang

To efficiently perform inference with neural networks, the underlying tensor programs require sufficient tuning efforts before being deployed into production environments. Usually, enormous tensor program candidates need to be sufficiently explored to find the one with the best performance. This is necessary to make the neural network products meet the high demand of real-world applications such as natural language processing, auto-driving, etc. Auto-schedulers are being developed to avoid the need for human intervention. However, due to the gigantic search space and lack of intelligent search guidance, current auto-schedulers require hours to days of tuning time to find the best-performing tensor program for the entire neural network. In this paper, we propose HARL, a reinforcement learning (RL) based auto-scheduler specifically designed for efficient tensor program exploration. HARL uses a hierarchical RL architecture in which learning-based decisions are made at all different levels of search granularity. It also automatically adjusts exploration configurations in real-time for faster performance convergence. As a result, HARL improves the tensor operator performance by 22% and the search speed by 4.3x compared to the state-of-the-art auto-scheduler. Inference performance and search speed are also significantly improved on end-to-end neural networks.

Zining Zhang, Yao Chen, Bingsheng He et al.

The increasing size and complexity of Large Language Models (LLMs) pose challenges for their deployment on personal computers and mobile devices. Aggressive post-training model compression is necessary to reduce the models' size, but it often results in significant accuracy loss. To address this challenge, we propose a novel network pruning technology that utilizes over 0.7 sparsity and less than 8 bits of quantization. Our approach enables the compression of prevailing LLMs within a couple of hours while maintaining a relatively small accuracy loss. In experimental evaluations, our method demonstrates effectiveness and potential for practical deployment. By making LLMs available on domestic devices, our work can facilitate a new era of natural language processing applications with wide-ranging impacts.

Han Wang, Alex Whitworth, Pak Ming Cheung et al.

Relevance evaluation plays a crucial role in personalized search systems to ensure that search results align with a user's queries and intent. While human annotation is the traditional method for relevance evaluation, its high cost and long turnaround time limit its scalability. In this work, we present our approach at Pinterest Search to automate relevance evaluation for online experiments using fine-tuned LLMs. We rigorously validate the alignment between LLM-generated judgments and human annotations, demonstrating that LLMs can provide reliable relevance measurement for experiments while greatly improving the evaluation efficiency. Leveraging LLM-based labeling further unlocks the opportunities to expand the query set, optimize sampling design, and efficiently assess a wider range of search experiences at scale. This approach leads to higher-quality relevance metrics and significantly reduces the Minimum Detectable Effect (MDE) in online experiment measurements.

Zining Zhang, Bingsheng He, Zhenjie Zhang

Speech separation is an important problem in speech processing, which targets to separate and generate clean speech from a mixed audio containing speech from different speakers. Empowered by the deep learning technologies over sequence-to-sequence domain, recent neural speech separation models are now capable of generating highly clean speech audios. To make these models more practical by reducing the model size and inference time while maintaining high separation quality, we propose a new transformer-based speech separation approach, called TransMask. By fully un-leashing the power of self-attention on long-term dependency exception, we demonstrate the size of TransMask is more than 60% smaller and the inference is more than 2 times faster than state-of-the-art solutions. TransMask fully utilizes the parallelism during inference, and achieves nearly linear inference time within reasonable input audio lengths. It also outperforms existing solutions on output speech audio quality, achieving SDR above 16 over Librimix benchmark.

Ruichu Cai, Jiawei Chen, Zijian Li et al.

Domain adaptation on time series data is an important but challenging task. Most of the existing works in this area are based on the learning of the domain-invariant representation of the data with the help of restrictions like MMD. However, such extraction of the domain-invariant representation is a non-trivial task for time series data, due to the complex dependence among the timestamps. In detail, in the fully dependent time series, a small change of the time lags or the offsets may lead to difficulty in the domain invariant extraction. Fortunately, the stability of the causality inspired us to explore the domain invariant structure of the data. To reduce the difficulty in the discovery of causal structure, we relax it to the sparse associative structure and propose a novel sparse associative structure alignment model for domain adaptation. First, we generate the segment set to exclude the obstacle of offsets. Second, the intra-variables and inter-variables sparse attention mechanisms are devised to extract associative structure time-series data with considering time lags. Finally, the associative structure alignment is used to guide the transfer of knowledge from the source domain to the target one. Experimental studies not only verify the good performance of our methods on three real-world datasets but also provide some insightful discoveries on the transferred knowledge.

Zining Zhang, Bingsheng He, Zhenjie Zhang

Non-parallel many-to-many voice conversion is recently attract-ing huge research efforts in the speech processing community. A voice conversion system transforms an utterance of a source speaker to another utterance of a target speaker by keeping the content in the original utterance and replacing by the vocal features from the target speaker. Existing solutions, e.g., StarGAN-VC2, present promising results, only when speech corpus of the engaged speakers is available during model training. AUTOVCis able to perform voice conversion on unseen speakers, but it needs an external pretrained speaker verification model. In this paper, we present our new GAN-based zero-shot voice conversion solution, called GAZEV, which targets to support unseen speakers on both source and target utterances. Our key technical contribution is the adoption of speaker embedding loss on top of the GAN framework, as well as adaptive instance normalization strategy, in order to address the limitations of speaker identity transfer in existing solutions. Our empirical evaluations demonstrate significant performance improvement on output speech quality and comparable speaker similarity to AUTOVC.

Yeqi Bai, Tao Ma, Lipo Wang et al.

While deep learning technologies are now capable of generating realistic images confusing humans, the research efforts are turning to the synthesis of images for more concrete and application-specific purposes. Facial image generation based on vocal characteristics from speech is one of such important yet challenging tasks. It is the key enabler to influential use cases of image generation, especially for business in public security and entertainment. Existing solutions to the problem of speech2face renders limited image quality and fails to preserve facial similarity due to the lack of quality dataset for training and appropriate integration of vocal features. In this paper, we investigate these key technical challenges and propose Speech Fusion to Face, or SF2F in short, attempting to address the issue of facial image quality and the poor connection between vocal feature domain and modern image generation models. By adopting new strategies on data model and training, we demonstrate dramatic performance boost over state-of-the-art solution, by doubling the recall of individual identity, and lifting the quality score from 15 to 19 based on the mutual information score with VGGFace classifier.

Zijian Li, Ruichu Cai, Kok Soon Chai et al.

Data-driven models are becoming essential parts in modern mechanical systems, commonly used to capture the behavior of various equipment and varying environmental characteristics. Despite the advantages of these data-driven models on excellent adaptivity to high dynamics and aging equipment, they are usually hungry to massive labels over historical data, mostly contributed by human engineers at an extremely high cost. The label demand is now the major limiting factor to modeling accuracy, hindering the fulfillment of visions for applications. Fortunately, domain adaptation enhances the model generalization by utilizing the labelled source data as well as the unlabelled target data and then we can reuse the model on different domains. However, the mainstream domain adaptation methods cannot achieve ideal performance on time series data, because most of them focus on static samples and even the existing time series domain adaptation methods ignore the properties of time series data, such as temporal causal mechanism. In this paper, we assume that causal mechanism is invariant and present our Causal Mechanism Transfer Network(CMTN) for time series domain adaptation. By capturing and transferring the dynamic and temporal causal mechanism of multivariate time series data and alleviating the time lags and different value ranges among different machines, CMTN allows the data-driven models to exploit existing data and labels from similar systems, such that the resulting model on a new system is highly reliable even with very limited data. We report our empirical results and lessons learned from two real-world case studies, on chiller plant energy optimization and boiler fault detection, which outperforms the existing state-of-the-art method.

Ruichu Cai, Jie Qiao, Kun Zhang et al.

Identification of causal direction between a causal-effect pair from observed data has recently attracted much attention. Various methods based on functional causal models have been proposed to solve this problem, by assuming the causal process satisfies some (structural) constraints and showing that the reverse direction violates such constraints. The nonlinear additive noise model has been demonstrated to be effective for this purpose, but the model class is not transitive--even if each direct causal relation follows this model, indirect causal influences, which result from omitted intermediate causal variables and are frequently encountered in practice, do not necessarily follow the model constraints; as a consequence, the nonlinear additive noise model may fail to correctly discover causal direction. In this work, we propose a cascade nonlinear additive noise model to represent such causal influences--each direct causal relation follows the nonlinear additive noise model but we observe only the initial cause and final effect. We further propose a method to estimate the model, including the unmeasured intermediate variables, from data, under the variational auto-encoder framework. Our theoretical results show that with our model, causal direction is identifiable under suitable technical conditions on the data generation process. Simulation results illustrate the power of the proposed method in identifying indirect causal relations across various settings, and experimental results on real data suggest that the proposed model and method greatly extend the applicability of causal discovery based on functional causal models in nonlinear cases.

Hoang Dung Vu, Kok Soon Chai, Bryan Keating et al.

Refrigeration and chiller optimization is an important and well studied topic in mechanical engineering, mostly taking advantage of physical models, designed on top of over-simplified assumptions, over the equipments. Conventional optimization techniques using physical models make decisions of online parameter tuning, based on very limited information of hardware specifications and external conditions, e.g., outdoor weather. In recent years, new generation of sensors is becoming essential part of new chiller plants, for the first time allowing the system administrators to continuously monitor the running status of all equipments in a timely and accurate way. The explosive growth of data flowing to databases, driven by the increasing analytical power by machine learning and data mining, unveils new possibilities of data-driven approaches for real-time chiller plant optimization. This paper presents our research and industrial experience on the adoption of data models and optimizations on chiller plant and discusses the lessons learnt from our practice on real world plants. Instead of employing complex machine learning models, we emphasize the incorporation of appropriate domain knowledge into data analysis tools, which turns out to be the key performance improver over state-of-the-art deep learning techniques by a significant margin. Our empirical evaluation on a real world chiller plant achieves savings by more than 7% on daily power consumption.

Ruichu Cai, Boyan Xu, Xiaoyan Yang et al.

Machine translation is going through a radical revolution, driven by the explosive development of deep learning techniques using Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). In this paper, we consider a special case in machine translation problems, targeting to convert natural language into Structured Query Language (SQL) for data retrieval over relational database. Although generic CNN and RNN learn the grammar structure of SQL when trained with sufficient samples, the accuracy and training efficiency of the model could be dramatically improved, when the translation model is deeply integrated with the grammar rules of SQL. We present a new encoder-decoder framework, with a suite of new approaches, including new semantic features fed into the encoder, grammar-aware states injected into the memory of decoder, as well as recursive state management for sub-queries. These techniques help the neural network better focus on understanding semantics of operations in natural language and save the efforts on SQL grammar learning. The empirical evaluation on real world database and queries show that our approach outperform state-of-the-art solution by a significant margin.

Ruichu Cai, Zhenjie Zhang, Zhifeng Hao

Causation discovery without manipulation is considered a crucial problem to a variety of applications. The state-of-the-art solutions are applicable only when large numbers of samples are available or the problem domain is sufficiently small. Motivated by the observations of the local sparsity properties on causal structures, we propose a general Split-and-Merge framework, named SADA, to enhance the scalability of a wide class of causation discovery algorithms. In SADA, the variables are partitioned into subsets, by finding causal cut on the sparse causal structure over the variables. By running mainstream causation discovery algorithms as basic causal solvers on the subproblems, complete causal structure can be reconstructed by combining the partial results. SADA benefits from the recursive division technique, since each small subproblem generates more accurate result under the same number of samples. We theoretically prove that SADA always reduces the scales of problems without sacrifice on accuracy, under the condition of local causal sparsity and reliable conditional independence tests. We also present sufficient condition to accuracy enhancement by SADA, even when the conditional independence tests are vulnerable. Extensive experiments on both simulated and real-world datasets verify the improvements on scalability and accuracy by applying SADA together with existing causation discovery algorithms.

Meihua Wang, Jiaming Mai, Yun Liang et al.

Traditional dehazing techniques, as a well studied topic in image processing, are now widely used to eliminate the haze effects from individual images. However, even the state-of-the-art dehazing algorithms may not provide sufficient support to video analytics, as a crucial pre-processing step for video-based decision making systems (e.g., robot navigation), due to the limitations of these algorithms on poor result coherence and low processing efficiency. This paper presents a new framework, particularly designed for video dehazing, to output coherent results in real time, with two novel techniques. Firstly, we decompose the dehazing algorithms into three generic components, namely transmission map estimator, atmospheric light estimator and haze-free image generator. They can be simultaneously processed by multiple threads in the distributed system, such that the processing efficiency is optimized by automatic CPU resource allocation based on the workloads. Secondly, a cross-frame normalization scheme is proposed to enhance the coherence among consecutive frames, by sharing the parameters of atmospheric light from consecutive frames in the distributed computation platform. The combination of these techniques enables our framework to generate highly consistent and accurate dehazing results in real-time, by using only 3 PCs connected by Ethernet.

Ganzhao Yuan, Yin Yang, Zhenjie Zhang et al.

Differential privacy enables organizations to collect accurate aggregates over sensitive data with strong, rigorous guarantees on individuals' privacy. Previous work has found that under differential privacy, computing multiple correlated aggregates as a batch, using an appropriate \emph{strategy}, may yield higher accuracy than computing each of them independently. However, finding the best strategy that maximizes result accuracy is non-trivial, as it involves solving a complex constrained optimization program that appears to be non-linear and non-convex. Hence, in the past much effort has been devoted in solving this non-convex optimization program. Existing approaches include various sophisticated heuristics and expensive numerical solutions. None of them, however, guarantees to find the optimal solution of this optimization problem. This paper points out that under ($ε$, $δ$)-differential privacy, the optimal solution of the above constrained optimization problem in search of a suitable strategy can be found, rather surprisingly, by solving a simple and elegant convex optimization program. Then, we propose an efficient algorithm based on Newton's method, which we prove to always converge to the optimal solution with linear global convergence rate and quadratic local convergence rate. Empirical evaluations demonstrate the accuracy and efficiency of the proposed solution.