Recent years have witnessed an explosive growth of AI models. The high cost of hosting AI services on GPUs and their demanding service requirements, make it timely and challenging to lower service costs and guarantee service quality. While spot instances have long been offered with a large discount, spot preemptions have discouraged users from using them to host model replicas when serving AI models.

Since pre-training and fine-tuning have been a successful paradigm in Natural Language Processing (NLP), this paper adopts the SOTA pre-training model-CeMAT as a strong assistant for low-resource ethnic language translation tasks. Aiming at the exposure bias problem in the fine-tuning process, we use the contrastive learning framework and propose a new contrastive examples generation method, which uses self- generated predictions as contrastive examples to expose the model to errors during inference. Moreover, in order to effectively utilize the limited bilingual data in low-resource tasks, this paper proposes a dynamic training strategy to fine-tune the model, and refines the model step by step by taking word embedding norm and uncertainty as the criteria of evaluate data and model respectively. Experimental results demonstrate that our method significantly improves the quality compared to the baselines, which fully verifies the effectiveness.

The Sky vision[3] aims to open a new era in cloud computing. Sky abstracts clouds and dynamically use multiple clouds to optimize workload execution. This enable users to focus on their business logic, rather than interact with multiple clouds, and manually optimize performance and costs.

Weak robustness and noise adaptability are major issues for Low-Resource Neural Machine Translation (NMT) models. Adversarial example is currently a major tool to improve model robustness and how to generate an adversarial examples that can degrade the performance of the model and ensure semantic consistency is a challenging task. In this paper, we adopt multi-reward reinforcement learning to generate adversarial examples for low-resource NMT. Specifically, utilizing gradient ascent to modify the source sentence, the discriminator and changes estimate are used to determine whether the generated adversarial examples maintain semantic consistency and the overall modifications of adversarial examples. Furthermore, we also install a language model reward to measure the fluency of adversarial examples. Experimental results on low-resource translation tasks show that our method highly aggressive to the model while maintaining semantic constraints greatly. Moreover, the model performance is significantly improved after fine-tuning with adversarial examples.

Multimodal Machine Translation (MMT) aims to enhance translation quality by incorporating information from other modalities (usually images). However, dominant MMT models do not consider that visual features not only provide supplementary information also introduce much noise. In this paper, we propose the visual features filter to solve this issue. Specifically, we adopt a soft-lookup function to select the visual features relevant to the text and then use these visual features as pseudo-words concatenating with a text representation. In addition, our model conducts two-pass decoding. The secondarypass decoding amounts to polishing which can identify errors in draft translations. The reason is that polishing expands the view in the process of decoding each target token, providing more contextual information. Besides, since most words in draft translations can be copied to final translations, we further equip our model with the copying mechanism to reserve those words that do not need to be corrected. MMT has achieved success in some mainstream languages at present. In order to promote the development of MMT in low-resource languages such as Mongolian, we deploy our model to the Mongolian→Chinese translation task. We expand Multi30k dataset to synthetic Mongolian and Chinese descriptions. Experiments on synthetic Mongolian and Chinese datasets demonstrate that our model can bring significant improvements.

Query optimizers are a performance-critical component in every database system. Due to their complexity, optimizers take experts months to write and years to refine. In this work, we demonstrate for the first time that learning to optimize queries without learning from an expert optimizer is both possible and efficient. We present Balsa, a query optimizer built by deep reinforcement learning. Balsa first learns basic knowledge from a simple, environment-agnostic simulator, followed by safe learning in real execution. On the Join Order Benchmark, Balsa matches the performance of two expert query optimizers, both open-source and commercial, with two hours of learning, and outperforms them by up to 2.8× in workload runtime after a few more hours. Balsa thus opens the possibility of automatically learning to optimize in future compute environments where expert-designed optimizers do not exist.

Deep autoregressive models compute point likelihood estimates of individual data points. However, many applications (i.e., database cardinality estimation), require estimating range densities, a capability that is under-explored by current neural density estimation literature. In these applications, fast and accurate range density estimates over high-dimensional data directly impact user-perceived performance. In this paper, we explore a technique for accelerating range density estimation over deep autoregressive models. This technique, called variable skipping, exploits the sparse structure of range density queries to avoid sampling unnecessary variables during approximate inference. We show that variable skipping provides 10-100x efficiency improvements when targeting challenging high-quantile error metrics, enables complex applications such as text pattern matching, and can be realized via a simple data augmentation procedure without changing the usual maximum likelihood objective.

Corporations today collect data at an unprecedented and accelerating scale, making the need to run queries on large datasets increasingly important. Technologies such as columnar block-based data organization and compression have become standard practice in most commercial database systems. However, the problem of best assigning records to data blocks on storage is still open. For example, today's systems usually partition data by arrival time into row groups, or range/hash partition the data based on selected fields. For a given workload, however, such techniques are unable to optimize for the important metric of the number of blocks accessed by a query. This metric directly relates to the I/O cost, and therefore performance, of most analytical queries. Further, they are unable to exploit additional available storage to drive this metric down further. In this paper, we propose a new framework called a query-data routing tree, or qd-tree, to address this problem, and propose two algorithms for their construction based on greedy and deep reinforcement learning techniques. Experiments over benchmark and real workloads show that a qd-tree can provide physical speedups of more than an order of magnitude compared to current blocking schemes, and can reach within 2X of the lower bound for data skipping based on selectivity, while providing complete semantic descriptions of created blocks.

Query optimizers rely on accurate cardinality estimates to produce good execution plans. Despite decades of research, existing cardinality estimators are inaccurate for complex queries, due to making lossy modeling assumptions and not capturing inter-table correlations. In this work, we show that it is possible to learn the correlations across all tables in a database without any independence assumptions. We present NeuroCard, a join cardinality estimator that builds a single neural density estimator over an entire database. Leveraging join sampling and modern deep autoregressive models, NeuroCard makes no inter-table or inter-column independence assumptions in its probabilistic modeling. NeuroCard achieves orders of magnitude higher accuracy than the best prior methods (a new state-of-the-art result of 8.5x maximum error on JOB-light), scales to dozens of tables, while being compact in space (several MBs) and efficient to construct or update (seconds to minutes).

Cardinality estimation has long been grounded in statistical tools for density estimation. To capture the rich multivariate distributions of relational tables, we propose the use of a new type of high-capacity statistical model: deep autoregressive models. However, direct application of these models leads to a limited estimator that is prohibitively expensive to evaluate for range or wildcard predicates. To produce a truly usable estimator, we develop a Monte Carlo integration scheme on top of autoregressive models that can efficiently handle range queries with dozens of dimensions or more. Like classical synopses, our estimator summarizes the data without supervision. Unlike previous solutions, we approximate the joint data distribution without any independence assumptions. Evaluated on real-world datasets and compared against real systems and dominant families of techniques, our estimator achieves single-digit multiplicative error at tail, an up to 90x accuracy improvement over the second best method, and is space- and runtime-efficient.

This paper describes Tacotron 2, a neural network architecture for speech synthesis directly from text. The system is composed of a recurrent sequence-to-sequence feature prediction network that maps character embeddings to mel-scale spectrograms, followed by a modified WaveNet model acting as a vocoder to synthesize time-domain waveforms from those spectrograms. Our model achieves a mean opinion score (MOS) of 4.53 comparable to a MOS of 4.58 for professionally recorded speech. To validate our design choices, we present ablation studies of key components of our system and evaluate the impact of using mel spectrograms as the conditioning input to WaveNet instead of linguistic, duration, and F0 features. We further show that using this compact acoustic intermediate representation allows for a significant reduction in the size of the WaveNet architecture.

A text-to-speech synthesis system typically consists of multiple stages, such as a text analysis frontend, an acoustic model and an audio synthesis module.Building these components often requires extensive domain expertise and may contain brittle design choices.In this paper, we present Tacotron, an end-to-end generative text-to-speech model that synthesizes speech directly from characters.Given pairs, the model can be trained completely from scratch with random initialization.We present several key techniques to make the sequence-tosequence framework perform well for this challenging task.Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English, outperforming a production parametric system in terms of naturalness.In addition, since Tacotron generates speech at the frame level, it's substantially faster than sample-level autoregressive methods.

We present ZipG, a distributed memory-efficient graph store for serving interactive graph queries. ZipG achieves memory efficiency by storing the input graph data using a compressed representation. What differentiates ZipG from other graph stores is its ability to execute a wide range of graph queries directly on this compressed representation. ZipG can thus execute a larger fraction of queries in main memory, achieving query interactivity. ZipG exposes a minimal API that is functionally rich enough to implement published functionalities from several industrial graph stores. We demonstrate this by implementing and evaluating graph queries from Facebook TAO, LinkBench, Graph Search and several other workloads on top of ZipG. On a single server with 244GB memory, ZipG executes tens of thousands of queries from these workloads for raw graph data over half a TB; this leads to an order of magnitude (sometimes as much as 23×) higher throughput than Neo4j and Titan. We get similar gains in distributed settings compared to Titan.

R is a popular statistical programming language with a number of extensions that support data processing and machine learning tasks. However, interactive data analysis in R is usually limited as the R runtime is single threaded and can only process data sets that fit in a single machine's memory. We present SparkR, an R package that provides a frontend to Apache Spark and uses Spark's distributed computation engine to enable large scale data analysis from the R shell. We describe the main design goals of SparkR, discuss how the high-level DataFrame API enables scalable computation and present some of the key details of our implementation.