Search

Energy harvesting creates an emerging intermittent computing paradigm but poses new challenges for sophisticated applications such as intermittent deep neural network (DNN) inference. Although model compression has adapted DNNs to resource-constrained devices, under intermittent power, compressed models will still experience multiple power failures during a single inference. Footprint-based approaches enable hardware-accelerated intermittent DNN inference by tracking footprints, independent of model computations, to indicate accelerator progress across power cycles. However, we observe that the extra overhead required to preserve progress indicators can severely offset the computation progress accumulated by intermittent DNN inference. This work proposes the concept of model augmentation to adapt DNNs to intermittent devices. Our middleware stack, JAPARI, appends extra neural network components into a given DNN, to enable the accelerator to intrinsically integrate progress indicators into the inference process, without affecting model accuracy. Their specific positions allow progress indicator preservation to be piggybacked onto output feature preservation to amortize the extra overhead, and their assigned values ensure uniquely distinguishable progress indicators for correct inference recovery upon power resumption. Evaluations on a Texas Instruments device under various DNN models, capacitor sizes, and progress preservation granularities show that JAPARI can speed up intermittent DNN inference by 3× over the state of the art, for common convolutional neural architectures that require heavy acceleration. [ABSTRACT FROM AUTHOR]

In this article, we pursue a better solution for the promising problem, i.e., the bidding strategy design, in the real-time bidding (RTB) advertising (AD) environment. Under the budget constraint, the design of an optimal strategy for bidding on each incoming impression opportunity targets at acquiring as many clicks as possible during an AD campaign. State-of-the-art bidding algorithms rely on a single predictor, the clickthrough rate predictor, to calculate the bidding value for each impression. This provides reasonable performance if the predictor has appropriate accuracy in predicting the probability of user clicking. However, the classical methods usually fail to capture optimal results since the predictor accuracy is limited. We improve the situation by accomplishing an additional winning price predictor in the bidding process. In this article, an algorithm combining powers of multiple prediction models is developed. It emerges froman analogy to the online stochastic knapsack problem, and the efficiency of the algorithm is also theoretically analyzed. Experiments conducted on real world RTB datasets show that the proposed solution performs better with regard to both number of clicks achieved and effective cost per click in many different settings of budget constraints. [ABSTRACT FROM AUTHOR]

When an event occurs, it attracts attention of information sources to publish related documents along its lifespan. The task of event detection is to automatically identify events and their related documents from a document stream, which is a set of chronologically ordered documents collected from various information sources. Generally, each event has a distinct activeness development so that its status changes continuously during its lifespan. When an event is active, there are a lot of related documents from various information sources. In contrast when it is inactive, there are very few documents, but they are focused. Previous works on event detection did not consider the characteristics of the event's activeness, and used rigid thresholds for event detection. We propose a concept called life profile, modeled by a hidden Markov model, to model the activeness trends of events. In addition, a general event detection framework, LIPED, which utilizes the learned life profiles and the burst-and-diverse characteristic to adjust the event detection thresholds adaptively, can be incorporated into existing event detection methods. Based on the official TDT corpus and contest rules, the evaluation results show that existing detection methods that incorporate LIPED achieve better performance in the cost and F1 metrics, than without. [ABSTRACT FROM AUTHOR]

In this article, we explore a novel sampling model, called feature preserved sampling (FPS) that sequentially generates a high-quality sample over sliding windows. The sampling quality we consider refers to the degree of consistency between the sample proportion and the population proportion of each attribute value in a window. Due to the time-variant nature of real-world datasets, users are more likely to be interested in the most recent data. However, previous works have not been able to generate a high-quality sample over sliding windows that precisely preserves upto- date population characteristics. Motivated by this shortcoming, we have developed the FPS algorithm, which has several advantages: (1) it sequentially generates a sample from a timevariant data source over sliding windows; (2) the execution time of FPS is linear with respect to the database size; (3) the relative proportional differences between the sample proportions and population proportions of most distinct attribute values are guaranteed to be below a specified error threshold, ϵ, while the relative proportion differences of the remaining attribute values are as close to ϵ as possible, which ensures that the generated sample is of high quality; (4) the sample rate is close to the user specified rate so that a high quality sampling result can be obtained without increasing the sample size; (5) by a thorough analytical and empirical study, we prove that FPS has acceptable space overheads, especially when the attribute values have Zipfian distributions, and FPS can also excellently preserve the population proportion of multivariate features in the sample; and (6) FPS can be applied to infinite streams and finite datasets equally, and the generated samples can be used for various applications. Our experiments on both real and synthetic data validate that FPS can effectively obtain a high quality sample of the desired size. In addition, while using the sample generated by FPS in various mining applications, a significant improvement in efficiency can be achieved without compromising the model's precision. [ABSTRACT FROM AUTHOR]

We investigate the general model of mining associations in a temporal database, where the exhibition periods of items are allowed to be different from one to another. The database is divided into partitions according to the time granularity imposed. Such temporal association rules allow us to observe short-term but interesting patterns that are absent when the whole range of the database is evaluated altogether. Prior work may omit some temporal association rules and thus have limited practicability. To remedy this and to give more precise frequent exhibition periods of frequent temporal itemsets, we devise an efficient algorithm Twain (standing for TWo end AssocIation miNer.) Twain not only generates frequent patterns with more precise frequent exhibition periods, but also discovers more interesting frequent patterns. Twain employs Start time and End time of each item to provide precise frequent exhibition period while progressively handling itemsets from one partition to another. Along with one scan of the database, Twain can generate frequent 2-itemsets directly according to the cumulative filtering threshold. Then, Twain adopts the scan reduction technique to generate all frequent k-itemsets (k > 2) from the generated frequent 2-itemsets. Theoretical properties of Twain are derived as well in this article. The experimental results show that Twain outperforms the prior works in the quality of frequent patterns, execution time, I/O cost, CPU overhead and scalability. [ABSTRACT FROM AUTHOR]