Search

Processing-in-memory (PIM) architectures show the advantage of handling applications that generate complicated memory request patterns; usually, those kinds of memory streams degrade the application’s performance in conventional memory hierarchy systems. In particular, deep convolutional neural networks (DCNNs) processing that consists of several functionalities could be highly optimized if PIM cores can extend the processing capability and data accessibility. In this work, we propose a functionality-based PIM accelerator for DCNNs. We design several modules in addition to the conventional PIM system based on a hybrid memory cube (HMC). First, we compose a new buffer module, namely, a shared cache, in which PIM cores are provided DCNN functionalities and pre-trained weights. The PIM cores subsequently enhance computational utilization and data accessibility. Second, an efficient replacement method complements the shared cache to optimize the data miss rate of DCNN processing. Third, we compose dual prefetchers that can deal with DCNN’s memory access patterns, thereby reducing the system’s overall latency. Fourth, we compose a PIM scheduler for PIM core-level autonomous request control. The PIM scheduler relieves the host processor of significant computational loads, achieving the overall latency of the system and reducing the energy consumption. By the performance evaluation based on the trace-driven HMC simulator, our proposed model improves average latency and bandwidth by 38.9 and 27.9 % with only 18.7 % more energy consumption compared with conventional HMC-based PIM systems. Our system also achieves scalable processing performance because when the DCNN becomes deeper, it processes faster than conventional PIM systems.

This paper presents an innovative prefetching algorithm for a hybrid main memory structure, which consists of DRAM and phase-change memory. To enhance the efficiency of hybrid memory hardware in serving big data technologies, the proposed design employs an application-adaptive algorithm based on big data execution characteristics. Specifically optimized for graph-processing applications, which exhibit complex and irregular memory access patterns, a dual prefetching scheme is proposed. This scheme comprises a fast-response model with low-cost algorithms for regular memory access patterns and an intelligent model based on an adaptive Gaussian-kernel-based machine-learning prefetch engine. The intelligent model can acquire knowledge from real-time data samples, capturing distinct memory access patterns via an adaptive Gaussian-kernel-based regression algorithm. These methods allow the model to self-adjust its hyperparameters at runtime, facilitating the implementation of locally weighted regression (LWR) for the Gaussian process of irregular access patterns. In addition, we introduced an efficient hybrid main memory architecture that integrates two different kinds of memory technologies, including DRAM and PCM, providing cost and energy efficiency over a DRAM-only memory structure. Based on the simulation-based experimental results, our proposed model achieved performance enhancement of 57% compared to the conventional DRAM model and of approximately 12% compared to existing prefetcher-based models.

The emergence of big data processing and machine learning has triggered the exponential growth of the working set sizes of applications. In addition, several modern applications are memory intensive with irregular memory access patterns. Therefore, we propose the concept of adaptive granularities to develop a prefetching methodology for analyzing memory access patterns based on a wider granularity concept that entails both cache lines and page granularity. The proposed prefetching module resides in the last-level cache (LLC) to handle the large working set of memory-intensive workloads. Additionally, to support memory access streams with variable intervals, we introduced an embedded-DRAM based LLC prefetch buffer that consists of three granularity-based prefetch engines and an access history table. By adaptively changing the granularity window for analyzing memory streams, the proposed model can swiftly and appropriately determine the stride of memory addresses to generate hidden delta chains from irregular memory access patterns. The proposed model achieves 18% and 15% improvements in terms of energy consumption and execution time compared to global history buffer and best offset prefetchers, respectively. In addition, our model reduced the total execution time and energy consumption by approximately 6% and 2.3%, compared to those of the Markov prefetcher and variable-length delta prefetcher.

The purpose of this study is to improve human emotional classification accuracy using a convolution neural networks (CNN) model and to suggest an overall method to classify emotion based on multimodal data. We improved classification performance by combining electroencephalogram (EEG) and galvanic skin response (GSR) signals. GSR signals are preprocessed using by the zero-crossing rate. Sufficient EEG feature extraction can be obtained through CNN. Therefore, we propose a suitable CNN model for feature extraction by tuning hyper parameters in convolution filters. The EEG signal is preprocessed prior to convolution by a wavelet transform while considering time and frequency simultaneously. We use a database for emotion analysis using the physiological signals open dataset to verify the proposed process, achieving 73.4% accuracy, showing significant performance improvement over the current best practice models.

Simultaneous localization and mapping (SLAM) is emerging as a prominent issue in computer vision and next-generation core technology for robots, autonomous navigation and augmented reality. In augmented reality applications, fast camera pose estimation and true scale are important. In this paper, we present an adaptive monocular visual–inertial SLAM method for real-time augmented reality applications in mobile devices. First, the SLAM system is implemented based on the visual–inertial odometry method that combines data from a mobile device camera and inertial measurement unit sensor. Second, we present an optical-flow-based fast visual odometry method for real-time camera pose estimation. Finally, an adaptive monocular visual–inertial SLAM is implemented by presenting an adaptive execution module that dynamically selects visual–inertial odometry or optical-flow-based fast visual odometry. Experimental results show that the average translation root-mean-square error of keyframe trajectory is approximately 0.0617 m with the EuRoC dataset. The average tracking time is reduced by 7.8%, 12.9%, and 18.8% when different level-set adaptive policies are applied. Moreover, we conducted experiments with real mobile device sensors, and the results demonstrate the effectiveness of performance improvement using the proposed method.

Human activity recognition is an emerging field of ubiquitous and pervasive computing. Although recent smartphones have powerful resources, the execution of machine learning algorithms on a large amount of data is still a burden on smartphones. Three major factors including; classification algorithm, data feature, and smartphone position influence the recognition accuracy and time. In this paper, we present a comparative study of six classification algorithms, six data features, and four different positions that are most commonly used in the recognition process using smartphone accelerometer. This analysis can be used to select any specific classification algorithm, data feature, and smartphone position for human activity recognition in terms of accuracy and response time. The methodology we used is composed of two major components; a data collector, and a classifier. A set of eleven activities of daily living, four different positions for data collection and ten volunteers contributed to make it a worth-full comparative study. Results show that K-Nearest Neighbor and J48 algorithms performed well both in terms of time and accuracy irrespective of data features whereas the performance of other algorithms is dependent on the selected data features. Similarly, mean and mode features gave good results in terms of accuracy irrespective of the classification algorithm. A short version of the paper has already been presented at ICIS 2014.

The essence of context awareness has changed the revolution of ubiquitous computing, and the wireless sensor network technologies paved the way towards many applications. Activity recognition is a key component in identifying the context of a user for providing services based on the application. In this study, we propose a context management model that is based on activity recognition. The model is composed of four components: a set of sensors, a set of activities, a backend server with machine learning algorithms, and a GUI application for the interaction with the user. A prototype is developed to show the usability of the proposed model. As a pilot testing, only accelerometer data of an Android phone is used to identify the activities of daily living (ADLs): sitting, standing, walking, and jogging. A good accuracy of results that is about 96% on average is achieved in all activities.