ABSTRACT
The spiking convolutional neural network (SCNN) is a kind of spiking neural network (SNN) with high accuracy for visual tasks and power efficiency on neuromorphic hardware, which is attractive for edge applications. However, it is challenging to implement SCNNs on resource-constrained edge devices because of the large number of convolutional operations and membrane potential (Vm) storage needed. Previous works have focused on timestep reduction, network pruning, and network quantization to realize SCNN implementation on edge devices. However, they overlooked similarities between spiking feature maps (SFmaps), which contain significant redundancy and cause unnecessary computation and storage. This work proposes a dual-threshold spiking convolutional neural network (DT-SCNN) to decrease the number of operations and memory access by utilizing similarities between SFmaps. The DT-SCNN employs dual firing thresholds to derive two similar SFmaps from one Vm map, reducing the number of convolutional operations and decreasing the volume of Vms and convolutional weights by half. We propose a variant spatio-temporal back propagation (STBP) training method with a two-stage strategy to train DT-SCNNs to decrease the inference timestep to 1. The experimental results show that the dual-thresholds mechanism achieves a 50% reduction in operations and data storage for the convolutional layers compared to conventional SCNNs while achieving not more than a 0.4% accuracy loss on the CIFAR10, MNIST, and Fashion MNIST datasets. Due to the lightweight network and single timestep inference, the DT-SCNN has the least number of operations compared to previous works, paving the way for low-latency and power-efficient edge applications.
ABSTRACT
Siamese networks have been extensively studied in recent years. Most of the previous research focuses on improving accuracy, while merely a few recognize the necessity of reducing parameter redundancy and computation load. Even less work has been done to optimize the runtime memory cost when designing networks, making the Siamese-network-based tracker difficult to deploy on edge devices. In this paper, we present SiamMixer, a lightweight and hardware-friendly visual object-tracking network. It uses patch-by-patch inference to reduce memory use in shallow layers, where each small image region is processed individually. It merges and globally encodes feature maps in deep layers to enhance accuracy. Benefiting from these techniques, SiamMixer demonstrates a comparable accuracy to other large trackers with only 286 kB parameters and 196 kB extra memory use for feature maps. Additionally, we verify the impact of various activation functions and replace all activation functions with ReLU in SiamMixer. This reduces the cost when deploying on mobile devices.
Subject(s)
Computers , Neural Networks, Computer , Computers, HandheldABSTRACT
Image demosaicking has been an essential and challenging problem among the most crucial steps of image processing behind image sensors. Due to the rapid development of intelligent processors based on deep learning, several demosaicking methods based on a convolutional neural network (CNN) have been proposed. However, it is difficult for their networks to run in real-time on edge computing devices with a large number of model parameters. This paper presents a compact demosaicking neural network based on the UNet++ structure. The network inserts densely connected layer blocks and adopts Gaussian smoothing layers instead of down-sampling operations before the backbone network. The densely connected blocks can extract mosaic image features efficiently by utilizing the correlation between feature maps. Furthermore, the block adopts depthwise separable convolutions to reduce the model parameters; the Gaussian smoothing layer can expand the receptive fields without down-sampling image size and discarding image information. The size constraints on the input and output images can also be relaxed, and the quality of demosaicked images is improved. Experiment results show that the proposed network can improve the running speed by 42% compared with the fastest CNN-based method and achieve comparable reconstruction quality as it on four mainstream datasets. Besides, when we carry out the inference processing on the demosaicked images on typical deep CNN networks, Mobilenet v1 and SSD, the accuracy can also achieve 85.83% (top 5) and 75.44% (mAP), which performs comparably to the existing methods. The proposed network has the highest computing efficiency and lowest parameter number through all methods, demonstrating that it is well suitable for applications on modern edge computing devices.
