Search

In this era of Transformers enjoying remarkable success, Convolutional Neural Networks (CNNs) remain highly relevant and useful. Indeed, hybrid Transformer-CNN network architectures, which combine the benefits of both approaches, have achieved impressive results. Vision Transformer (ViT) is a significant neural network architecture that features a convolutional layer as its first layer, primarily built on the transformer framework. However, owing to the distinct computation patterns inherent in attention and convolution, existing hardware accelerators for these two models are typically designed separately and lack a unified approach toward accelerating both models efficiently. In this paper, we present a dedicated accelerator on a field-programmable gate array (FPGA) platform. The accelerator, which integrates a configurable three-dimensional systolic array, is specifically designed to accelerate the inferential capabilities of hybrid Transformer-CNN networks. The Convolution and Transformer computations can be mapped to a systolic array by unifying these operations for matrix multiplication. Softmax and LayerNorm which are frequently used in hybrid Transformer-CNN networks were also implemented on FPGA boards. The accelerator achieved high performance with a peak throughput of 722 GOP/s at an average energy efficiency of 53 GOPS/W. Its respective computation latencies were 51.3 ms, 18.1 ms, and 6.8 ms for ViT-Base, ViT-Small, and ViT-Tiny. The accelerator provided a 12× improvement in energy efficiency compared to the CPU, a 2.3× improvement compared to the GPU, and a 1.5× to 2× improvement compared to existing accelerators regarding speed and energy efficiency.

To achieve real-time object detection tasks with high throughput and low latency, this paper proposes a multi-die hardware accelerator architecture. It implements three accelerators on the VU9P chip, each of which is bound to an independent super logic region (SLR). To reduce off-chip memory access and power consumption, this design uses three on-chip buffers to store the weights and intermediate result data on one hand; on the other hand, it minimizes data access and movement and maximizes data reuse. This design uses an 8-bit quantization strategy for both weights and feature maps to achieve twice the throughput and computational efficiency of a single digital signal processor (DSP). In addition, many operators are designed in the accelerator, and all of them are fully parameterized, so it is easy to extend the network, and the control of the accelerator can be realized by configuring the instruction group. By accelerating the YOLOv4-tiny algorithm, the accelerator architecture can achieve a frame rate of 148.14 frames per second (FPS) and a peak throughput of 2.76 tera operations per second (TOPS) at 200 MHz with an energy efficiency ratio of 93.15 GOPS/W. The code can be found at https://github.com/19801201/Verilog_CNN_Accelerator.

This article proposed a novel classification framework that can classify the samples of multiple domains based on the outputs of multiple models. Different from the existing methods that train single model on all domains, our framework trains multiple models on each domain. On a testing sample, the outputs of all trained models are used to predict the domain of this sample. Then, this sample is classified by the output of models that belong to the predicted domain. Experiments show that our framework achieved higher accuracy than the existing methods. Furthermore, our framework achieves good scalability on multiple domains.

Generally, the performance of deep learning-based classification models is highly related to the captured features of training samples. When a sample is not clear or contains a similar number of features of many objects, we cannot easily classify what it is. Actually, human beings classify objects by not only the features but also some information such as the probability of these objects in an environment. For example, when we know further information such as one object has a higher probability in the environment than the others, we can easily give the answer about what is in the sample. We call this kind of probability as local probability as this is related to the local environment. In this paper, we carried out a new framework that is named L-PDL to improve the performance of deep learning based on the analysis of this kind of local probability. Firstly, our method trains the deep learning model on the training set. Then, we can get the probability of objects on each sample by this trained model. Secondly, we get the posterior local probability of objects on the validation set. Finally, this probability conditionally cooperates with the probability of objects on testing samples. We select three popular deep learning models on three real datasets for the evaluation. The experimental results show that our method can obviously improve the performance on the real datasets, which is better than the state-of-the-art methods.

When utilizing the deep learning models in some real applications, the distribution of the labels in the environment can be used to increase the accuracy. Generally, to compute this distribution, there should be the validation set that is labeled by the ground truths. On the other side, the dependency of ground truths limits the utilization of the distribution in various environments. In this paper, we carried out a novel system for the deep learning-based classification to solve this problem. Firstly, our system only uses one validation set with ground truths to compute some hyper parameters, which is named as one-shot guidance. Secondly, in an environment, our system builds the validation set and labels this by the prediction results, which does not need any guidance by the ground truths. Thirdly, the computed distribution of labels by the validation set selectively cooperates with the probability of labels by the output of models, which is to increase the accuracy of predict results on testing samples. We selected six popular deep learning models on three real datasets for the evaluation. The experimental results show that our system can achieve higher accuracy than state-of-art methods while reducing the dependency of labeled validation set. [ABSTRACT FROM AUTHOR]

Objective Systematic reviews and meta-analyses have revealed the associations between H. pylori infection and various health outcomes. We aimed to evaluate the strength and breadth of evidence on the associations. Design Umbrella review of systematic reviews and meta-analyses. Setting No settings. Participants No patients involved. Data sources Embase, PubMed, Web of Science, Cochrane Library Databases, CNKI, VIP database and Wangfang database from inception to February 1, 2019. Outcomes measures Diverse diseases (such as cancer and ischaemic heart disease). Results Sixty articles reporting 88 unique outcomes met the eligible criteria. 74 unique outcomes had nominal significance (p<0.05). Of the outcomes with significance, 61 had harmful associations and 13 had beneficial associations. Furthermore, 73% (64) of the outcomes exhibited significant heterogeneity . Of the these meta-analyses, 32 had moderate to high heterogeneity (I²=50%–75%) and 24 had high heterogeneity (I²>75%). Moreover, 20% exhibited publication bias (p<0.1). In addition, 97% of the methodological qualities were rated ‘critically low’. 36% of the evidence qualities of outcomes were rated ‘low’, 56% of the evidence qualities were rated ‘very low’ and 8% of the evidence qualities were rated ‘moderate’. H. pylori infection may be associated with an increased risk of five diseases and a decreased risk of irritable bowel syndrome. Conclusion Although 60 meta-analyses explored 88 unique outcomes, moderate quality evidence only existed for six outcomes with statistical significance. H. pylori infection may be associated with a decreased risk of irritable bowel syndrome and an increased risk of hypertriglyceridemia, chronic cholecystitis and cholelithiasis, gestational diabetes mellitus, gastric cancer and systemic sclerosis. [ABSTRACT FROM AUTHOR]