Abstract In this paper, we propose a pipeline to investigate the performance of semantic segmentation model that employs an encoder-decoder architecture with atrous separable convolution and spatial pyramid pooling, trained on multi-resolution whole slide breast pathological images with different patch sizes. Our segmentation model obtains the best performance on zoom level 2 (10 $$\times $$ × magnification) with AUC score 0.974 in terms of slide-level classification. This outperforms both the performance of the pathologist and other semantic segmentation models on the Camelyon16 dataset. By offering a larger field of view and reducing noise and detail, training a semantic segmentation model on the properly selected lower resolution pathology images can further improve the precision of pixel-wise cancer region segmentation. By contrast, the corresponding inference time is 14 times shorter than the inference time trained on the highest resolution patches, and it is also shorter than the time required by a pathologist with time constraints. Moreover, we prove that the model trained on lower resolution patches can still generate refined external polygons of cancer region on the highest resolution image. This study provides new insights into efficient gigapixel histopathology analysis that will make clinical adoption more likely.

Traditional datacenter analysis is based on high-level, coarse-grained metrics. This obscures our vision of datacenter behavior, as we do not observe the full picture nor subtleties that might make up these high-level, coarse metrics. There is room for operational improvement based on fine-grained temporal and spatial, low-level metric data. We leverage in this work one of the (rare) public datasets providing fine-grained information on datacenter operations, with over 60 billion measurements captured in 15-second intervals. We show evidence that fine-grained information reveals new operational aspects, that the different metrics cannot be derived from one another (and thus need to be captured), and that many low-level metrics, gathered frequently are key to understanding datacenter operations. We propose a holistic analysis for datacenter operations, providing statistical characterization of node and workload aspects. Our analysis reveals both generic and machine learning-specific aspects, summarized in over 30 observations, providing deep insight into this dataset and the originating cluster. We give actionable insights, surprising findings, and exemplify how our observations support performance-engineering tasks such as workload prediction and long-term datacenter design.

Deep learning is finding its way into high energy physics by replacing traditional Monte Carlo simulations. However, deep learning still requires an excessive amount of computational resources. A promising approach to make deep learning more efficient is to quantize the parameters of the neural networks to reduced precision. Reduced precision computing is extensively used in modern deep learning and results to lower execution inference time, smaller memory footprint and less memory bandwidth. In this paper we analyse the effects of low precision inference on a complex deep generative adversarial network model. The use case which we are addressing is calorimeter detector simulations of subatomic particle interactions in accelerator based high energy physics. We employ the novel Intel low precision optimization tool (iLoT) for quantization and compare the results to the quantized model from TensorFlow Lite. In the performance benchmark we gain a speed-up of 1.73x on Intel hardware for the quantized iLoT model compared to the initial, not quantized, model. With different physics-inspired self-developed metrics, we validate that the quantized iLoT model shows a lower loss of physical accuracy in comparison to the TensorFlow Lite model.

In this paper we propose a novel supervised learning approach for training Artificial Neural Networks (ANNs) to evaluate chess positions. The method that we present aims to train different ANN architectures to understand chess positions similarly to how highly rated human players do. We investigate the capabilities that ANNs have when it comes to pattern recognition, an ability that distinguishes chess grandmasters from more amateur players. We collect around 3,000,000 different chess positions played by highly skilled chess players and label them with the evaluation function of Stockfish, one of the strongest existing chess engines. We create 4 different datasets from scratch that are used for different classification and regression experiments. The results show how relatively simple Multilayer Perceptrons (MLPs) outperform Convolutional Neural Networks (CNNs) in all the experiments that we have performed. We also investigate two different board representations, the first one representing if a piece is present on the board or not, and the second one in which we assign a numerical value to the piece according to its strength. Our results show how the latter input representation influences the performances of the ANNs negatively in almost all experiments.

Small to medium enterprises (SMEs), particularly those whose business is focused on developing innovative produces, are limited by a major bottleneck in the speed of computation in many applications. The recent developments in GPUs have been the marked increase in their versatility in many computational areas. But due to the lack of specialist GPUprogramming skills, the explosion of GPU power has not been fully utilized in general SME applications by inexperienced users. Also, the existing automatic CPU-to-GPU code translators are mainly designed for research purposes with poor user interface design and are hard to use. Little attentions have been paid to the applicability, usability, and learnability of these tools for normal users. In this paper, we present an online automated CPU-to-GPU source translation system (GPSME) for inexperienced users to utilize the GPU capability in accelerating general SME applications. This system designs and implements a directive programming model with a new kernel generation scheme and memory management hierarchy to optimize its performance. A web service interface is designed for inexperienced users to easily and flexibly invoke the automatic resource translator. Our experiments with nonexpert GPU users in four SMEs reflect that a GPSME system can efficiently accelerate real-world applications with at least 4× and have a better applicability, usability, and learnability than the existing automatic CPU-to-GPU source translators.

Summary The support vector machine (SVM) is a supervised learning algorithm used for recognizing patterns in data. It is a very popular technique in machine learning and has been successfully used in applications such as image classification, protein classification, and handwriting recognition. However, the computational complexity of the kernelized version of the algorithm grows quadratically with the number of training examples. To tackle this high computational complexity, we have developed a directive‐based approach that converts a gradient‐ascent based training algorithm for the CPU to an efficient graphics processing unit (GPU) implementation. We compare our GPU‐based SVM training algorithm to the standard LibSVM CPU implementation, a highly optimized GPU‐LibSVM implementation, as well as to a directive‐based OpenACC implementation. The results on different handwritten digit classification datasets demonstrate an important speed‐up for the current approach when compared to the CPU and OpenACC versions. Furthermore, our solution is almost as fast and sometimes even faster than the highly optimized CUBLAS‐based GPU‐LibSVM implementation, without sacrificing the algorithm's accuracy. Copyright © 2014 John Wiley & Sons, Ltd.

Visualizing very large graphs by edge bundling is a promising method, yet subject to several challenges: speed, clutter, level-of-detail, and parameter control. We present CUBu, a framework that addresses the above problems in an integrated way. Fully GPU-based, CUBu bundles graphs of up to a million edges at interactive framerates, being over 50 times faster than comparable state-of-the-art methods, and has a simple and intuitive control of bundling parameters. CUBu extends and unifies existing bundling techniques, offering ways to control bundle shapes, separate bundles by edge direction, and shade bundles to create a level-of-detail visualization that shows both the graph core structure and its details. We demonstrate CUBu on several large graphs extracted from real-life application domains.