Search

Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware. This paper evaluates a custom ASIC---called a Tensor Processing Unit (TPU) --- deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X -- 30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X -- 80X higher. Moreover, using the CPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

Predictive analytics play an important role in clinical research. An accurate predictive model can help clinicians stratify risk thereby allowing the identification of a target population which might benefit from a certain intervention. Conventionally, predictive analytics is performed using parametric modeling which comes with a number of assumptions. For example, generalized linear regression models require linearity and additivity to hold for the underlying data. However, these assumptions may not hold in practice. Especially in the era of big data, a large number of covariates or features can be extracted from an electronic database which might have complex interactions and higher-order terms among the covariates. Conventional modeling methods have trouble capturing such high-dimensional relationships. However, some sophisticated machine learning techniques have been invented to handle this situation. Gradient boosting is one of these techniques which is able to recursively fit a weak learner to the residual so as to improve model performance with a gradually increasing number of iterations. It can automatically discover complex data structure, including nonlinearity and high-order interactions, even in the context of hundreds, thousands, or tens-of-thousands of potential predictors. This paper aims to introduce how gradient boosting works. The principles behind this learning machine are explained with a small example in a step-by-step manner. The formal implementation of gradient tree boosting is then illustrated with the caret package. In the simulated example complexity of data structure is created by generating certain interactions between the covariates. This example shows that gradient boosting can better capture these complex relationships than a generalized linear model-based approach.

Today’s credit card issuers are increasingly offering a broad range of products and services with separate lines of business responsible for different product groups. Too often, the separate lines of business operate independently and information available to one line of business may not be used productively by others. In this study, we examine the potential of using information from customers of multiple products to identify customers most likely to respond to cross-sell product offers. Specifically, we examine the potential for offering home loans to a population of credit card holders by studying individuals who do hold both a credit card and a mortgage with the card issuer. Using real world data provided to the 2007 PAKDD data mining competition, we employ Friedman’s stochastic gradient boosting (MART™, TreeNet® ) for the rapid development of a high performance cross-sell predictive model.

This paper presents the top 10 data mining algorithms identified by the IEEE International Conference on Data Mining (ICDM) in December 2006: C4.5, k-Means, SVM, Apriori, EM, PageRank, AdaBoost, kNN, Naive Bayes, and CART. These top 10 algorithms are among the most influential data mining algorithms in the research community. With each algorithm, we provide a description of the algorithm, discuss the impact of the algorithm, and review current and further research on the algorithm. These 10 algorithms cover classification, clustering, statistical learning, association analysis, and link mining, which are all among the most important topics in data mining research and development.

Mobile phone customers face many choices regarding handset hardware, add-on services, and features to subscribe to from their service providers. Mobile phone companies are now increas-ingly interested in the drivers of migration to third generation (3G) hardware and services. Using real world data provided to the 10th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) 2006 Data Mining Competition we explore the effectiveness of Friedman’s stochastic gradient boosting (Multiple Additive Regression Trees [MART]) for the rapid development of a high performance predictive model.

This paper introduces new forms, sampling and estimation approaches fordiscrete choice models. The new models include behavioral specifications oflatent class choice models, multinomial probit, hybrid logit, andnon-parametric methods. Recent contributions also include new specializedchoice based sample designs that permit greater efficiency in datacollection. Finally, the paper describes recent developments in the use ofsimulation methods for model estimation. These developments are designed toallow the applications of discrete choice models to a wider variety ofdiscrete choice problems.

Experimental choice analysis continues to attract academic and applied attention. We review what is known about the design, conduct, analysis, and use of data from choice experiments, and indicate gaps in current knowledge that should be addressed in future research. Design strategies consistent with probabilistic models of choice process and the parallels between choice experiments and real markets are considered. Additionally, we address the issues of reliability and validity. Progress has been made in accounting for differences in reliability, but more research is needed to determine which experiments and response procedures will consistently produce more reliable data for various problems.

The past few years have witnessed a growing recognition of soft computing technologies that underlie the conception, design and utilization of intelligent systems. According to Zadeh [1], soft computing consists of artificial neural networks, fuzzy inference system, approximate reasoning and derivative free optimization techniques. In this paper, we report a performance analysis among Multivariate Adaptive Regression Splines (MARS), neural networks and neuro-fuzzy systems. The MARS procedure builds flexible regression models by fitting separate splines to distinct intervals of the predictor variables. For performance evaluation purposes, we consider the famous Box and Jenkins gas furnace time series benchmark. Simulation results show that MARS is a promising regression technique compared to other soft computing techniques.

Long-term rainfall prediction is a challenging task especially in the modern world where we are facing the major environmental problem of global warming. In general, climate and rainfall are highly non-linear phenomena in nature exhibiting what is known as the "butterfly effect". While some regions of the world are noticing a systematic decrease in annual rainfall, others notice increases in flooding and severe storms. The global nature of this phenomenon is very complicated and requires sophisticated computer modeling and simulation to predict accurately. In this paper, we report a performance analysis for Multivariate Adaptive Regression Splines (MARS) [1] and artificial neural networks for one month ahead prediction of rainfall. To evaluate the prediction efficiency, we made use of 87 years of rainfall data in Kerala state, the southern part of the Indian peninsula situated at latitude-longitude pairs (8°29 N - 76°57 E). We used an artificial neural network trained using the scaled conjugate gradient algorithm. The neural network and MARS were trained with 40 years of rainfall data. For performance evaluation, network predicted outputs were compared with the actual rainfall data. Simulation results reveal that MARS is a good forecasting tool and performed better than the considered neural network.