Shallow SqueezeNext: An Efficient & Shallow DNN

In recent years, Convolution Neural Network (CNN) gained great success in many applications, especially in computer vision. Now adapting CNN inference on edge devices has become an active research area in embedded vision and hot topic in Edge AI. The primary design hurdles for implementing CNN inference on embedded systems are limited computational resource, memory resource, and power budget. This study presents a novel architecture for SqueezeNet [1] like CNN models, which can be extended to support any CNN model as well. We address two approaches to mitigate resource constraints. First, we use a custom floating point(12 bit for computation and 8bit for storing). Second is slicing the model into a repetitive block called computation blocks. Computation block can be configured dynamically by the host processor to operate in a different mode. We have implemented SqueezeNet v1.1 for Image-Net [2] for large-scale classification which achieved around 9 FPS at 100MHz. SqueezeNet, mapped for our architecture achieves top-1 accuracy of 51% for the ImageNet dataset. Unlike other implementations which use FPGA boards with a large amount of resources, our experiments are done in DE10 Nano which mimics actual embedded system like environment.

Wide spread of IoT devices has made it possible to acquire enormous amounts of realtime sensor information. Due to the explosive increase in the sensing data volume, it becomes difficult to collect and process all the data in one central place. On one hand, storing and processing data on edge devices, so called edge computing, is becoming important. On the other hand, edge devices usually have only limited computing and memory resources, and hence it is not practical to process and save all the acquired data. There is a great demand of effectively selecting data to process on an edge device or to transfer it to a cloud server. In this paper, we propose an efficient density-based data selection and management method called O-D2M by which edge devices store the data representing inherent data distribution. We use a low cost graph algorithm to analyze input data trend and its density. We evaluate effectiveness of the proposed O-D2M comparing to other methods in terms of the accuracy of machine learning models trained by the selected data. Throughout the evaluation, we confirm that O-D2M obtains higher accuracy and lower computation cost while it can reduce the amount of data to be processed or transferred by up to 20 points.

The advancement of the Internet of Things has enabled numerous applications ranging from smart wearable to connected vehicles. However, the limited energy and memory resources of low-end IoT devices significantly impede their further flourish. In this article, we consider a system that consists of an IoT device and an edge server. The edge server stores code blocks for the IoT device and loads required blocks to the IoT device for execution thereby alleviates the latter from the limited memory resource. Furthermore, the IoT device can harvest energy from the ambient energy sources to achieve a sustainable operation. To deal with the dynamic energy harvesting process and block request process, we propose a stochastic block prefetching framework (BPF) to optimize the user experienced delay. The BPF assists the IoT device to intelligently prefetch blocks from the edge server according to the historical user behaviors. The BPF consists of three modules, i.e., estimation module, prefetching module, and dual learning module. The estimation module measures the probability of block being requested in the future. The prefetching module requests blocks from the edge server according to the available energy and memory. The dual learning module helps to accelerate the convergence of the framework. The numerous simulation results are provided to verify the effectiveness of the proposed framework.

The use of the Karhunen-Loeve Transform (KLT) for spectral decorrelation in compression of hyperspectral satellite images results in improved performance. However, the KLT algorithm consists of sequential processes, which are computationally intensive, such as the Covariance and Eigenvector evaluations, etc. These processes slow down the overall computation of the KLT transform significantly. The acceleration of these processes within the context of limited power and hardware budgets is the main objective of this paper. The computations of each of these processes are investigated thoroughly by breaking them down into primitive arithmetic operations. Subsequently, a comprehensive analysis of these computations is presented to inspect the possibility and feasibility of different acceleration techniques, such as parallelism. The proposed designs are implemented on a System-on-a-Chip platform, which incorporates a 32-bit hardwired microcontroller and a coprocessing unit built within a field programmable gate array fabric. Two novel architectures are proposed offering accelerated processing within a very limited power budget (less than 0.225 Watt). The proposed solution is not only feasible for space applications, but also for different mobile and remote sensing applications.

In recent years, with the rise of large language models, model sizes have grown dramatically, garnering attention for their remarkable performance but also raising concerns about the substantial computational and communication resources they require. This has created significant challenges in fine-tuning or re-training models on devices with limited computing and memory resources. Efficient model compression through low-rank factorization has emerged as a promising solution, offering a way to balance the tradeoff between compression ratio and prediction accuracy. However, existing approaches to low-rank selection often rely on trial-and-error methods to determine the optimal rank, lacking theoretical guidance and incurring high computational costs. Furthermore, these methods typically treat low-rank factorization as a post-training process, resulting in suboptimal compressed models. In this paper, we design a novel approach by integrating rank selection into the low-rank training process and performing independent layer-wise rank selection under the guidance of a theoretical loss error bound. Specifically, we first conduct a comprehensive theoretical analysis to quantify how low-rank approximations impact the training losses. Building on these insights, we develop an efficient layer-wise rank search algorithm and seamlessly incorporate it into low-rank singular value decomposition (SVD) training. Our evaluation results on benchmark datasets demonstrate that our approach can achieve high prediction accuracy while delivering significant compression performance. Furthermore, our solution is generic and can be extended to broader learning models.

CrossPrune: Cooperative pruning for camera–LiDAR fused perception models of autonomous driving

As more and more cloud services are exposed to DDoS attacks, DDoS attack detect has become a new challenging task because large packet traces captured on fast links could not be easily handled on a single server with limited computing and memory resources. In this paper, we propose a Hadoop based model to identify abnormal packets and compute the statistics according to the number of abnormal packets. The novelties of the model are that:(1) by harnessing HBASE, an improved bloom filter based mapping mechanism named TCP2HC/UDP2HC are implemented, (2)with the characteristics of IP spoofing and the temporal correlation of transport layer connection state, an extensible set of rules and a reliable MapReduce based checking mechanism for abnormal packets are designed, (3) using statistic features extracted from the increased abnormal packets and TCP/UDP flow, a non-parameter CUSUM algorithm is used to detect most DDoS attacks accurately and efficiently. The model can detect the attack behavior in the early stage, which is beneficial to mitigate attack with the help of flow cleaning by converting a check rule to the filtering rule. Experiments show no matter how large the attack scale and what kind of DDoS attack, the detection model can soon detect DDoS attack accurately.

The concept of compressing deep Convolutional Neural Networks (CNNs) is essential to use limited computation, power, and memory resources on embedded devices. However, existing methods achieve this objective at the cost of a drop in inference accuracy in computer vision tasks. To address such a drawback, we propose a framework that leverages knowledge distillation along with customizable block-wise optimization to learn a lightweight CNN structure while preserving better control over the compression-performance tradeoff. Considering specific resource constraints, e.g., floating-point operations per inference (FLOPs) or model-parameters, our method results in a state of the art network compression while being capable of achieving better inference accuracy. In a comprehensive evaluation, we demonstrate that our method is effective, robust, and consistent with results over a variety of network architectures and datasets, at negligible training overhead. In particular, for the already compact network MobileNet v2, our method offers up to 2× and 5.2× better model compression in terms of FLOPs and model-parameters, respectively, while getting 1.05% better model performance than the baseline network.

The substitution box (S-box) is one of the most important components in modern ciphers. This paper reports a new class of S-boxes defined by the logarithmic functions in GF(16). Due to their small sizes, these S-boxes can be used to construct a compact cipher for the devices with limited computing and memory resources, such as those in wireless sensor or mobile communications.

Model compression is crucial for the deployment of neural networks on devices with limited computational and memory resources. Many different methods show comparable accuracy of the compressed model and similar compression rates. However, the majority of the compression methods are based on heuristics and offer no worst case guarantees on the tradeoff between the compression rate and the approximation error for an arbitrarily new sample. We propose the first efficient structured pruning algorithm with a provable tradeoff between its compression rate and the approximation error for any future test sample. Our method is based on the coreset framework, and it approximates the output of a layer of neurons/filters by a coreset of neurons/filters in the previous layer and discards the rest. We apply this framework in a layer-by-layer fashion from the bottom to the top. Unlike previous works, our coreset is data-independent, meaning that it provably guarantees the accuracy of the function for any input [Formula: see text], including an adversarial one.

Modern convolutional neural networks (CNNs) in computer vision are trained on a large number of images from numerous categories to form rich discriminative feature extractors. Inference using such models on resource-constrained Internet-of-Things (IoT) platforms poses a challenge and an opportunity. Having limited computation, storage, and energy budgets, most IoT platforms are not capable of hosting such compute intensive models. However, typical IoT applications demand detection of a relatively small number of categories, albeit the specific categories of interest may change at runtime as the context evolves dynamically. In this letter, we take advantage of the opportunity to address the challenge. Specifically, we develop a novel transformation to the architecture of a given CNN, so that the majority of the inference workload is allocated to class-specific disjoint branches, which can be dynamically executed or skipped, based on the context, to fulfill the application requirements. Experiments demonstrate that our approach preserves the classification accuracy for the classes of interest, while proportionally decreasing the model complexity and inference workload.

Pose estimation has been a hot topic in the field of machine vision in recent years. Animals exist widely in nature, and the analysis of their shape and movement is important in many fields and industries. In the pose estimation task, to improve the detection accuracy, the existing models often need to consume a lot of computing and memory resources. Therefore, it is a key problem for the pose estimation methods to carry out a lightweight model and reduce the computational overhead on the premise of ensuring model accuracy. In this paper, we focus on the structure of the convolutional neural network in animal pose estimation, construct a lightweight and efficient stacked hourglass network model oriented to optimize the balance of model computation and accuracy, and implement the application algorithm design based on it. Aiming at the problem of large parameters in depthwise convolutional neural networks, a lightweight residual module is proposed, that is, based on the lightweight efficient channel attention improved conditional channel-weighted method (ICCW-Bottle), thereby reducing the weight of the network and obtaining the feature information of different scales. Given the problem that a large amount of feature information is easily lost after the network pooling operation, a lightweight dual-branch fusion module is proposed that fully integrates high-level semantic information and low-level detailed features under the condition of a small number of parameters. Finally, the same as the CC-SSL method: the model is trained jointly using synthetic and real animal datasets, but the CC-SSL method does not take into account the computational power of the model, which consumes a lot of time and memory to run. Through experiments, it is known that compared with the CC-SSL method, the PCK@0.05 of this method is increased by 5.5% on the TigDog dataset. The model in this paper reduces the number of parameters and calculations of the network while ensuring less information loss and model accuracy. The ablation experiment verifies the advancement and effectiveness of the overall network.

Wireless Sensor Networks (WSN) comprise autonomous embedded nodes that combine sensing, actuation, and distributed computing with small size and low energy. Testing of WSNs is challenging due to distributed functionality, lack of test and debug interfaces, and limited communication, computation and memory resources. This paper presents the design and implementation of a practical network monitoring and analysis tool for identifying the causes of misbehavior. The tool consists of a sniffer node that passively captures WSN traffic, and multiple user interfaces that can run e.g. on a PC or mobile phone. Unlike the related proposals, our tool neither needs setting up an additional monitoring network alongside the actual WSN nor uses any node resources. Practical testing experiences with mul-tihop mesh WSN deployments show that the tool reveals design and implementation defects that are hard to discover with other testing methods such as in-network monitoring systems or debug prints.

Applications such as Simultaneous Localization and Mapping (SLAM) can greatly benefit from RGB-D sensor data to produce 3D maps of the environment as well as sensor’s trajectory estimation. However, the resulting 3D points map can be cumbersome, and since indoor environments are mainly composed of planar surfaces, the idea is to use planes as building blocks for a SLAM process. This paper describes an RGB-D SLAM system benefiting from planes segmentation to generate lightweight 3D plane-based maps. Our goal is to produce reduced 3D maps composed solely of planes sections that can be used on platforms with limited memory and computation resources. We present the introduction of planar regions in a regular RGB-D SLAM system and evaluate the benefits regarding both resulting map and estimated camera trajectory.

Heart sounds classification plays an important role in cardiovascular disease detection. Currently, deep learning methods for heart sound classification with heavy parameters consumption cannot be deployed in environments with limited memory and computational budgets. Besides, de-noising of heart sound signals (HSSs) can affect accuracy of heart sound classification, because erroneous removal of meaningful components may lead to heart sound distortion. In this paper, an automated heart sound classification method using adaptive wavelet threshold and 1D LDCNN (One-dimensional Lightweight Deep Convolutional Neural Network) is proposed. In this method, we exploit WT (Wavelet Transform) with an adaptive threshold to de-noise heart sound signals (HSSs). Furthermore, we utilize 1D LDCNN to realize automatic feature extraction and classification for de-noised heart sounds. Experiments on PhysioNet/CinC 2016 show that our proposed method achieves the superior classification results and excels in consumption of parameter comparing to state-of-the-art methods.