Reducing the amount of annotated data required to train predictive models is one of the main challenges in applying artificial intelligence to histopathology. In this paper, we propose a method to enhance the performance of deep learning models trained with limited data in the field of digital pathology. The method relies on a two-stage transfer learning process, where an intermediate model serves as a bridge between a pre-trained model on ImageNet and the final cancer classification model. The intermediate model is fine-tuned with a dataset of over 4,000,000 images weakly labeled with clinical data extracted from TCGA program. The model obtained through the proposed method significantly outperforms a model trained with a traditional transfer learning process.
Kota Padang merupakan kota budaya dan terletak pada lintasan geografis yang sangat strategis. Potensi wisata Kota Padang yang begitu besar belum dapat dioptimalkan dengan baik karena minimnya media promosi yang digunakan oleh pemerintah. Pariwisata sebagai salah satu kegiatan yang komprehensif, memerlukan suatu sistem yang dapat menyajikan informasi mengenai lokasi wisata dan informasi penunjang kepariwisataan seperti infomasi peta lokasi dan rute menuju lokasi objek wisata tersebut. Aplikasi WebGis dipilih karena penyajian informasinya lebih menarik dan sangat menggambarkan kondisi sebenarnya yang ditampilkan dalam peta, baik peta garis maupun model permukaan digital. Aplikasi Sistem Informasi Geografis Pariwisata Kota Padang yang dikembangkan dengan kerangka website HTML, dengan bahasa pemograman (JavaScript dan PHP), MySql sebagai basis data dan Google Maps Api. Hasil akhir dari penelitian ini adalah terciptanya aplikasi WebGis yang dapat menyajikan informasi pariwisata kota padang yang terintegrasi, terinci dan dapat di akses dengan mudah.
In a 1961 lecture to celebrate MIT's centennial, John McCarthy proposed the vision of utility computing, including three key concepts of pay-per-use service, large computer and private computer. Six decades have passed, but Mc- Carthy's computing utility vision has not yet been fully realized, despite advances in grid computing, services computing and cloud computing. This paper presents a perspective of computing utility called Information Superbahn, building on recent advances in cloud computing. This Information Superbahn perspective retains McCarthy's vision as much as possible, while making essential modern requirements more explicit, in the new context of a networked world of billions of users, trillions of devices, and zettabytes of data. Computing utility offers pay-per-use computing services through a 1) planet-scale, 2) low-entropy and 3) high-goodput utility. The three salient characteristics of computing utility are elaborated. Initial evidence is provided to support this viewpoint.
In the era of the Internet of Things, Bluetooth low energy (BLE/BTLE) plays an important role as a well-known wireless communication technology. While the security and privacy of BLE have been analyzed and fixed several times, the threat of side-channel attacks to BLE devices is still not well understood. In this work, we highlight a side-channel threat to the re-keying protocol of BLE. This protocol uses a fixed long term key for generating session keys, and the leakage of the long term key could render the encryption of all the following (and previous) connections useless. Our attack exploits the side-channel leakage of the re-keying protocol when it is implemented on embedded devices. In particular, we present successful correlation electromagnetic analysis and deep learning based profiled analysis that recover long term keys of BLE devices. We evaluate our attack on an ARM Cortex-M4 processor (Nordic Semiconductor nRF52840) running Nimble, a popular open-source BLE stack. Our results demonstrate that the long term key can be recovered within only a small amount of electromagnetic traces. Further, we summarize the features and limitations of our attack, and suggest a range of countermeasures to prevent it.
Current Retinex-based image enhancement methods with fixed scale filters cannot adapt to situations involving various depths of field and illuminations. In this paper, a simple but effective method based on adaptive full-scale Retinex (AFSR) is proposed to clarify underwater images or videos. First, we design an adaptive full-scale filter that is guided by the optical transmission rate to estimate illumination components. Then, to reduce the computational complexity, we develop a quantitative mapping method instead of non-linear log functions for directly calculating the reflection component. The proposed method is capable of real-time processing of underwater videos using temporal coherence and Fourier transformations. Compared with eight state-of-the-art clarification methods, our method yields comparable or better results for image contrast enhancement, color-cast correction and clarity.
Exploring the expected quantizing scheme with suitable mixed-precision policy is the key to compress deep neural networks (DNNs) in high efficiency and accuracy. This exploration implies heavy workloads for domain experts, and an automatic compression method is needed. However, the huge search space of the automatic method introduces plenty of computing budgets that make the automatic process challenging to be applied in real scenarios. In this paper, we propose an end-to-end framework named AutoQNN, for automatically quantizing different layers utilizing different schemes and bitwidths without any human labor. AutoQNN can seek desirable quantizing schemes and mixed-precision policies for mainstream DNN models efficiently by involving three techniques: quantizing scheme search (QSS), quantizing precision learning (QPL), and quantized architecture generation (QAG). QSS introduces five quantizing schemes and defines three new schemes as a candidate set for scheme search, and then uses the Differentiable Neural Architecture Search (DNAS) algorithm to seek the layer- or model-desired scheme from the set. QPL is the first method to learn mixed-precision policies by reparameterizing the bitwidths of quantizing schemes, to the best of our knowledge. QPL optimizes both classification loss and precision loss of DNNs efficiently and obtains the relatively optimal mixed-precision model within limited model size and memory footprint. QAG is designed to convert arbitrary architectures into corresponding quantized ones without manual intervention, to facilitate end-to-end neural network quantization. We have implemented AutoQNN and integrated it into Keras. Extensive experiments demonstrate that AutoQNN can consistently outperform state-of-the-art quantization. For 2-bit weight and activation of AlexNet and ResNet18, AutoQNN can achieve the accuracy results of 59.75% and 68.86%, respectively, and obtain accuracy improvements by up to 1.65% and 1.74%, respectively, compared with state-of-the-art methods. Especially, compared with the full-precision AlexNet and ResNet18, the 2-bit models only slightly incur accuracy degradation by 0.26% and 0.76%, respectively, which can fulfill practical application demands.
Image matting is to estimate the opacity of foreground objects from an image. A few deep learning based methods have been proposed for image matting and perform well in capturing spatially close information. However, these methods fail to capture global contextual information, which has been proved essential in improving matting performance. This is because a matting image may be up to several megapixels, which is too big for a learning-based network to capture global contextual information due to the limit size of a receptive field. Although uniformly downsampling the matting image can alleviate this problem, it may result in the degradation of matting performance. To solve this problem, we introduce a natural image matting with the attended global context method to extract global contextual information from the whole image, and to condense them into a suitable size for learning-based network. Specifically, we first leverage a deformable sampling layer to obtain condensed foreground and background attended images respectively. Then, we utilize a contextual attention layer to extract information related to unknown regions from condensed foreground and background images generated by a deformable sampling layer. Besides, our network predicts a background as well as the alpha matte to obtain more purified foreground, which contributes to better qualitative performance in composition. Comprehensive experiments show that our method achieves competitive performance on both Composition-1k and the alphamatting.com benchmark quantitatively and qualitatively.
Identifying semantic types for attributes in relations, known as attribute semantic type (AST) identification, plays an important role in many data analysis tasks, such as data cleaning, schema matching, and keyword search in databases. However, due to a lack of unified naming standards across prevalent information systems (a.k.a. information islands), AST identification still remains as an open problem. To tackle this problem, we propose a context-aware method to figure out the ASTs for relations in this paper. We transform the AST identification into a multi-class classification problem and propose a schema context aware (SCA) model to learn the representation from a collection of relations associated with attribute values and schema context. Based on the learned representation, we predict the AST for a given attribute from an underlying relation, wherein the predicted AST is mapped to one of the labeled ASTs. To improve the performance for AST identification, especially for the case that the predicted semantic types of attributes are not included in the labeled ASTs, we then introduce knowledge base embeddings (a.k.a. KBVec) to enhance the above representation and construct a schema context aware model with knowledge base enhanced (SCA-KB) to get a stable and robust model. Extensive experiments based on real datasets demonstrate that our context-aware method outperforms the state-of-the-art approaches by a large margin, up to 6.14% and 25.17% in terms of macro average F 1 score, and up to 0.28% and 9.56% in terms of weighted F 1 score over high-quality and low-quality datasets respectively.
Agile hardware development methodology has been widely adopted over the past decade. Despite the research progress, the industry still doubts its applicability, especially for the functional verification of complicated processor chips. Functional verification commonly employs a simulation-based method of co-simulating the design under test with a reference model and checking the consistency of their outcomes given the same input stimuli. We observe limited collaboration and information exchange through the design and verification processes, dramatically leading to inefficiencies when applying the conventional functional verification workflow to agile development. In this paper, we propose workflow integration with collaborative task delegation and dynamic information exchange as the design principles to effectively address the challenges on functional verification under the agile development model. Based on workflow integration, we enhance the functional verification workflows with a series of novel methodologies and toolchains. The diff-rule based agile verification methodology (DRAV) reduces the overhead of building reference models with runtime execution information from designs under test. We present the RISC-V implementation for DRAV, DiffTest, which adopts information probes to extract internal design behaviors for co-simulation and debugging. It further integrates two plugins, namely XFUZZ for effective test generation guided by design coverage metrics and LightSSS for efficient fault analysis triggered by co-simulation mismatches. We present the integrated workflows for agile hardware development and demonstrate their effectiveness in designing and verifying RISC-V processors with 33 functional bugs found in N ut S hell . We also illustrate the efficiency of the proposed toolchains with a case study on a functional bug in the L2 cache of X iang S han .
On the Internet, we find a large amount of information from government institutions that has been published in open format. However, only a part of these data is available in standard formats such as Resource Description Framework (RDF), and to a lesser extent, is published as Linked Open Data (LOD). The main objective of the research presented in this paper is to identify problems and tools used in the process of publishing LOD with the purpose of establishing a basis for the construction of a future framework that will help public institutions to facilitate such processes. To fulfill the objective, we conducted a systematic literature review in order to assess the state-of-the-art in this matter. The contribution of this work is to identify the frequent problems that arise in the LOD publishing process. It also provides a detail of the frameworks proposed in scientific papers grouping the technical tools by phases that correspond to the LOD publication life cycle. In addition, it compiles the characteristics of the ETL (Extract-Transform-Load) tools that predominate in this review, such as Pentaho Data Integration (Kettle) and OpenRefine.
Building an effective sequential recommendation system is still a challenging task due to limited interactions among users and items. Recent work has shown the effectiveness of incorporating textual or visual information into sequential recommendation to alleviate the data sparse problem. The data sparse problem now is attracting a lot of attention in both industry and academic community. However, considering interactions among modalities on a sequential scenario is an interesting yet challenging task because of multimodal heterogeneity. In this paper, we introduce a novel recommendation approach of considering both textual and visual information, namely Multimodal Interactive Network (MIN). The advantage of MIN lies in designing a learning framework to leverage the interactions among modalities from both the item level and the sequence level for building an efficient system. Firstly, an item-wise interactive layer based on the encoder-decoder mechanism is utilized to model the item-level interactions among modalities to select the informative information. Secondly, a sequence interactive layer based on the attention strategy is designed to capture the sequence-level preference of each modality. MIN seamlessly incorporates interactions among modalities from both the item level and the sequence level for sequential recommendation. It is the first time that interactions in each modality have been explicitly discussed and utilized in sequential recommenders. Experimental results on four real-world datasets show that our approach can significantly outperform all the baselines in sequential recommendation task.
Dynamic neural network (NN) techniques are increasingly important because they facilitate deep learning techniques with more complex network architectures. However, existing studies, which predominantly optimize the static computational graphs by static scheduling methods, usually focus on optimizing static neural networks in deep neural network (DNN) accelerators. We analyze the execution process of dynamic neural networks and observe that dynamic features introduce challenges for efficient scheduling and pipelining in existing DNN accelerators. We propose DyPipe, a holistic approach to optimizing dynamic neural network inferences in enhanced DNN accelerators. DyPipe achieves significant performance improvements for dynamic neural networks while it introduces negligible overhead for static neural networks. Our evaluation demonstrates that DyPipe achieves 1.7x speedup on dynamic neural networks and maintains more than 96% performance for static neural networks.
Due to the unprecedented development of low-latency interconnect technology, building large-scale disaggregated architecture is drawing more and more attention from both industry and academia. Resource disaggregation is a new way to organize the hardware resources of datacenters, and has the potential to overcome the limitations, e.g., low resource utilization and low reliability, of conventional datacenters. However, the emerging disaggregated architecture brings severe performance and latency problems to the existing cloud systems. In this paper, we take memory disaggregation as an example to demonstrate the unique challenges that the disaggregated datacenter poses to the existing cloud software stacks, e.g., programming interface, language runtime, and operating system, and further discuss the possible ways to reinvent the cloud systems.
Technological advances in recent years have promoted the development of virtual reality systems that have a wide variety of hardware and software characteristics, providing varying degrees of immersion. Immersion is an objective property of the virtual reality system that depends on both its hardware and software characteristics. Virtual reality systems are currently attempting to improve immersion as much as possible. However, there is no metric to measure the level of immersion of a virtual reality system based on its characteristics. To date, the influence of these hardware and software variables on immersion has only been considered individually or in small groups. The way these system variables simultaneously affect immersion has not been analyzed either. In this paper, we propose immersion metrics for virtual reality systems based on their hardware and software variables, as well as the development process that led to their formulation. From the conducted experiment and the obtained data, we followed a methodology to generate immersion models based on the variables of the system. The immersion metrics presented in this work offer a useful tool in the area of virtual reality and immersive technologies, not only to measure the immersion of any virtual reality system but also to analyze the relationship and importance of the variables of these systems.
Emerging technologies, e.g., cloud computing, artificial intelligence, robotics, and information technologies, facilitate digital transformation and promote digital services. Significant research exists on the acceptance and use of technologies and specific digital transformation (DT) practices. However, more research is needed to better understand user’s acceptance and how they use technology to better de-sign digital services. Thus, this paper: 1) proposes a Systematic Literature Re-view (SLR) of the models of acceptance and use of technologies, as well as of certain DT practices applied to the human-computer interaction (HCI) of digital services; 2) provides an analysis of the results of the behavior of the factors and practices; and 3) proposes a taxonomy of the acceptance and use factors applied to HCI of digital services by economic activity. The main contribution of this work is the proposed taxonomy in a vacancy area relevant to digital transformation efforts applied to digital services pursued globally, both in the private and public sectors.
Prevalent use of motion capture (MoCap) produces large volumes of data and MoCap data retrieval becomes crucial for efficient data reuse. MoCap clips may not be neatly segmented and labeled, increasing the difficulty of retrieval. In order to effectively retrieve such data, we propose an elastic content-based retrieval scheme via unsupervised posture encoding and strided temporal alignment (PESTA) in this work. It retrieves similarities at the sub-sequence level, achieves robustness against singular frames and enables control of tradeoff between precision and efficiency. It firstly learns a dictionary of encoded postures utilizing unsupervised adversarial autoencoder techniques and, based on which, compactly symbolizes any MoCap sequence. Secondly, it conducts strided temporal alignment to align a query sequence to repository sequences to retrieve the best-matching sub-sequences from the repository. Further, it extends to find matches for multiple sub-queries in a long query at sharply promoted efficiency and minutely sacrificed precision. Outstanding performance of the proposed scheme is well demonstrated by experiments on two public MoCap datasets and one MoCap dataset captured by ourselves.
To overcome the limitations of long latency and privacy concerns from cloud computing, edge computing along with distributed machine learning such as federated learning (FL), has gained much attention and popularity in academia and industry. Most existing work on FL over the edge mainly focuses on optimizing the training of one shared global model in edge systems. However, with the increasing applications of FL in edge systems, there could be multiple FL models from different applications concurrently being trained in the shared edge cloud. Such concurrent training of these FL models can lead to edge resource competition (for both computing and network resources), and further affect the FL training performance of each other. Therefore, in this paper, considering a multi-model FL scenario, we formulate a joint participant selection and learning optimization problem in a shared edge cloud. This joint optimization aims to determine FL participants and the learning schedule for each FL model such that the total training cost of all FL models in the edge cloud is minimized. We propose a multi-stage optimization framework by decoupling the original problem into two or three subproblems that can be solved respectively and iteratively. Extensive evaluation has been conducted with realworld FL datasets and models. The results have shown that our proposed algorithms can reduce the total cost efficiently compared with prior algorithms.
The evaluation of human-machine interfaces (HMI) requires quantitative metrics to define the ability of a person to effectively achieve their goals using the HMI. In particular, for pointing-device type HMIs such as the computer mouse, an experiment quantifying movement by performing repetitive target selections allows defining a useful metric known as throughput (TP) using the Fitts’ Law test. In this work, a dataset obtained from an automated protocol application is presented, which is made publicly available through an on-line platform. A post-processing method to obtain performance parameters from the dataset is also presented, and its output is used to validate the data against similar experiments in the literature.
Computer vision (CV) algorithms have been extensively used for a myriad of applications nowadays. As the multimedia data are generally well-formatted and regular, it is beneficial to leverage the massive parallel processing power of the underlying platform to improve the performances of CV algorithms. Single Instruction Multiple Data (SIMD) instructions, capable of conducting the same operation on multiple data items in a single instruction, are extensively employed to improve the efficiency of CV algorithms. In this paper, we evaluate the power and effectiveness of RISC-V vector extension (RV-V) on typical CV algorithms, such as Gray Scale, Mean Filter, and Edge Detection. By our examinations, we show that compared with the baseline OpenCV implementation using scalar instructions, the equivalent implementations using the RV-V (version 0.8) can reduce the instruction count of the same CV algorithm up to 24x, when processing the same input images. Whereas, the actual performances improvement measured by the cycle counts is highly related with the specific implementation of the underlying RV-V co-processor. In our evaluation, by using the vector co-processor (with eight execution lanes) of Xuantie C906, vector-version CV algorithms averagely exhibit up to 2.98x performances speedups compared with their scalar counterparts.
This paper focuses on end-to-end task-oriented dialogue systems,which jointly handle dialogue state track-ing(DST)and response generation.Traditional methods usually adopt a supervised paradigm to learn DST from a manu-ally labeled corpus.However,the annotation of the corpus is costly,time-consuming,and cannot cover a wide range of do-mains in the real world.To solve this problem,we propose a multi-span prediction network(MSPN)that performs unsu-pervised DST for end-to-end task-oriented dialogue.Specifically,MSPN contains a novel split-merge copy mechanism that captures long-term dependencies in dialogues to automatically extract multiple text spans as keywords.Based on these keywords,MSPN uses a semantic distance based clustering approach to obtain the values of each slot.In addition,we pro-pose an ontology-based reinforcement learning approach,which employs the values of each slot to train MSPN to generate relevant values.Experimental results on single-domain and multi-domain task-oriented dialogue datasets show that MSPN achieves state-of-the-art performance with significant improvements.Besides,we construct a new Chinese dialogue dataset MeDial in the low-resource medical domain,which further demonstrates the adaptability of MSPN.
