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.

Improved tuberculosis (TB) prevention and control depend critically on the development of a simple, readily accessible rapid triage test to stratify TB risk. We hypothesized that a blood protein-based host response signature for active TB (ATB) could distinguish it from other TB-like disease (OTD) in adult patients with persistent cough, thereby providing a foundation for a point-of-care (POC) triage test for ATB. Three adult cohorts consisting of ATB suspects were recruited. A bead-based immunoassay and machine learning algorithms identified a panel of four host blood proteins, interleukin-6 (IL-6), IL-8, IL-18, and vascular endothelial growth factor (VEGF), that distinguished ATB from OTD. An ultrasensitive POC-amenable single-molecule array (Simoa) panel was configured, and the ATB diagnostic algorithm underwent blind validation in an independent, multinational cohort in which ATB was distinguished from OTD with receiver operator characteristic-area under the curve (ROC-AUC) of 0.80 [95% confidence interval (CI), 0.75 to 0.85], 80% sensitivity (95% CI, 73 to 85%), and 65% specificity (95% CI, 57 to 71%). When host antibodies against TB antigen Ag85B were added to the panel, performance improved to 86% sensitivity and 69% specificity. A blood-based host response panel consisting of four proteins and antibodies to one TB antigen can help to differentiate ATB from other causes of persistent cough in patients with and without HIV infection from Africa, Asia, and South America. Performance characteristics approach World Health Organization (WHO) target product profile accuracy requirements and may provide the foundation for an urgently needed blood-based POC TB triage test.

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.

Developing load participation in ancillary services to the electric grid. Challenges: coordination among multiple entities, targeted R&D for market conditions and regulatory and policy environments.

We show that the binary logistic regression model can often be estimated even when the study sample is confined to observations on only one of the possible outcomes of the dependent variable. Provided that an appropriate supplementary sample can be found, the two samples may be pooled, and a simple method employed to estimate the model with a conventional statistics package. The supplementary sample must contain information on the regressors of the model, but need not contain any information on the dependent variable. Hence supplementary samples can often be found in general purpose public use surveys such as the U.S. Census.

The 1984 monograph, “CART: Classification and Regression Trees,” coauthored by Leo Breiman, Jerome Friedman, Richard Olshen, and Charles Stone (BFOS), represents a major milestone in the evolution of artificial intelligence, machine learning, nonparametric statistics, and data mining. The work is important for the comprehensiveness of its study of decision trees, the technical innovations it introduces, its sophisticated examples of tree-structured data analysis, and its authoritative treatment of large sample theory for trees. Since its publication the CART monograph has been cited some 3000 times according to the science and social science citation indexes; Google Scholar reports about 8,450 citations. CART citations can be found in almost any domain, with many appearing in fields such as credit risk, targeted marketing, financial markets modeling, electrical engineering, quality control, biology, chemistry, and clinical medical research. CART has also strongly influenced image compressionvia tree-structured vector quantization. This brief account is intended to introduce CART basics, touching on the major themes treated in the CART monograph, and to encourage readers to return to the rich original source for technical details, discussions revealing the thought processes of the authors, and examples of their analytical style.

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.

A new method for estimating the demand curve for publicly supplied goods when quantities are restricted to a few discrete levels is introduced. The method involves fitting a conditional logit model to choices from a set of survey options in which price and quantity are both varied and consumer attitudes are explicitly controlled. The estimated parameters of the valuation function serve to trace the marginal value of the good at each level of hypothetical consumption in survey data. We apply the method to the valuation of salmon on Alaska's Kenai River. We find that there is a distinct kink in the marginal valuation function and that sport fishermen may place a negative marginal value on fish permits exceeding their desired catch levels.

In the late 1970's, a welfare reform experiment with 'workfare'-like provisions was conducted in Minnesota. The experiment provided education and training to AFDC recipients in an attempt to improve client long-term labor-market outcomes. An evaluation of the experiment applying simultaneous-equations modeling of outcomes, attrition, and endogenously missing data is presented. We find imprecise evidence that the program increased the probability of post-program work participation, and we measure a high rate of return in wages to work experience. Although we find no evidence of attrition bias, there is considerable influence of selectivity in the wage equations.