Optimization of Neural Network Computation with use of Residual Number System for Tasks of Design of Neural Network Systems of Automatic Control

Tikhonov, E.E.; Sosin, A.I.; Evdokimov, A. A.

doi:10.1109/fareastcon.2018.8602679

Cited by 10 publications

(2 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…If the Boolean value is zero, it indicates that the packet should not be forwarded from the corresponding port, otherwise, it indicates that the flow should be forwarded along with the port as shown in Table 2. Generally, NDN data flows are thus able to de-map the same RPI from the above calculation process, a more detailed description of the CRT algorithm can be found in [15], and also some papers focused on optimal RNS can be found in [20], [21], and [22].…”

Section: Querying Group Membershipmentioning

confidence: 99%

An RNS-Based NDN Forwarding Plane: Design and Evaluation

Jia

2023

IEEE Access

View full text Add to dashboard Cite

Due to the variable length and uniqueness of names in NDN, traditional IP address-based routing and forwarding architecture is no longer suitable for NDN networks. It is still a challenge to design a high-performance forwarding engine to solve the problem of variable name length and uniqueness in NDN. In this article, Residue Number System (RNS) is applied to the routing and forwarding of NDN for the first time, and a new forwarding engine framework based on the RNS is designed. The simulation results show that this method has a significant performance increase in memory space and system latency. At the core-level node, the RNS-based process requires only about 5 percent of Bloom Filter (BF) in terms of memory space and has a 9.8 percent decrease in system latency compared to applying BF (P=1E-08). This method not only achieves less memory usage and reduces the system latency but also fundamentally solves the false positives and has good scalability due to the superiority of the algorithm itself. The proposed work is admirably suited for realizing large-scale NDNs by hardware implementation of NDN routers.INDEX TERMS Named data networking (NDN), packet forwarding engine (PFE), residue number system (RNS), membership querying.

show abstract

Section: Querying Group Membershipmentioning

confidence: 99%

An RNS-Based NDN Forwarding Plane: Design and Evaluation

Jia

2023

IEEE Access

View full text Add to dashboard Cite

show abstract

“…Investigating different methods for performing convolution operations, such as 1x1 [1], spatially separable [2], depth-wise separable [3], and shuffled grouped convolutions [4]. Proposing techniques to reduce the number of operations such as pruning [5], quantization [6], numbering representation [7], and fast algorithms [8]. Exploring new architectures based on dataflow [9].…”

Section: Introductionmentioning

confidence: 99%

SLID: Exploiting Spatial Locality in Input Data as a Computational Reuse Method for Efficient CNN

et al. 2021

View full text Add to dashboard Cite

Convolutional Neural Networks (CNNs) revolutionized computer vision and reached the state-of-the-art performance for image processing, object recognition, and video classification. Even though CNN inference is notoriously compute-intensive, as convolutions account for > 90% of the total operation tasks, the ability to tradeoff between accuracy, performance, power, and latency to meet target application makes it an open research topic. This paper proposes the Spatial Locality Input Data (SLID) method for computational reuse during the inference stage for a pre-trained network. The method exploits input data spatial locality via skipping partial processing of the multiply-and-accumulate (MAC) operations for adjacent data and equating its value to previously computed ones. SLID improves the throughput of resource-constrained devices (Internet-of-Things, edge devices) and accelerates computations during the inference phase by reducing the number of MAC operations. Such approximate computing schema does not require a similarity quantification step nor any modification for the training stage. The computational data reuse was evaluated on three well-known distinctive CNN structures and data sets with alternating layer selections: LeNet, CIFAR-10, and AlexNet. The computational data reuse method saves up to 34.9%, 49.84%, and 31.5% of MAC operations while reducing the accuracy by 8%, 3.7%, and 5.0% for the three models mentioned earlier, respectively. Besides, the proposed method saves on memory access by eliminating data fetching of skipped inputs. Furthermore, filter size, strides, and padding on the accuracy and savings of operations are analyzed. SLID is the first work to exploit the input spatial locality for savings on CNN convolution operations with minimal accuracy loss and without memory or computational overhead. This makes it a great option to support intelligence at the edge.

show abstract