Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020551
Meiqi Wang, Jianqiao Mo, Jun Lin, Zhongfeng Wang, L. Du
Early-exit is a kind of technique to terminate a pre-specified computation at an early stage depending on the input samples and has been introduced to reduce energy consumption for Deep Neural Networks (DNNs). Previous early-exit approaches suffered from the burden of manually tuning early-exit loss-weights to find a good trade-off between complexity reduction and system accuracy. In this work, we first propose DynExit, a dynamic loss-weight modification strategy for ResNets, which adaptively modifies the ratio of different exit branches and searches for a proper spot for both accuracy and cost. Then, an efficient hardware unit for early-exit branches is developed, which can be easily integrated to existing hardware architectures of DNNs to reduce average computing latency and energy cost. Experimental results show that the proposed DynExit strategy can reduce up to 43.6% FLOPS compared to the state-of-the-arts approaches. On the other hand, it is able to achieve 1.2% accuracy improvement over the existing end-to-end fixed loss-weight training scheme with comparable computation reduction ratio. The proposed hardware architecture for DynExit is evaluated on the platform of Xilinx Zynq-7000 ZC706 development board. Synthesis results demonstrate that the architecture can achieve high speed with low hardware complexity. To the best of our knowledge, this is the first hardware implementation for early-exit techniques used for DNNs in open literature.
{"title":"DynExit: A Dynamic Early-Exit Strategy for Deep Residual Networks","authors":"Meiqi Wang, Jianqiao Mo, Jun Lin, Zhongfeng Wang, L. Du","doi":"10.1109/SiPS47522.2019.9020551","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020551","url":null,"abstract":"Early-exit is a kind of technique to terminate a pre-specified computation at an early stage depending on the input samples and has been introduced to reduce energy consumption for Deep Neural Networks (DNNs). Previous early-exit approaches suffered from the burden of manually tuning early-exit loss-weights to find a good trade-off between complexity reduction and system accuracy. In this work, we first propose DynExit, a dynamic loss-weight modification strategy for ResNets, which adaptively modifies the ratio of different exit branches and searches for a proper spot for both accuracy and cost. Then, an efficient hardware unit for early-exit branches is developed, which can be easily integrated to existing hardware architectures of DNNs to reduce average computing latency and energy cost. Experimental results show that the proposed DynExit strategy can reduce up to 43.6% FLOPS compared to the state-of-the-arts approaches. On the other hand, it is able to achieve 1.2% accuracy improvement over the existing end-to-end fixed loss-weight training scheme with comparable computation reduction ratio. The proposed hardware architecture for DynExit is evaluated on the platform of Xilinx Zynq-7000 ZC706 development board. Synthesis results demonstrate that the architecture can achieve high speed with low hardware complexity. To the best of our knowledge, this is the first hardware implementation for early-exit techniques used for DNNs in open literature.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"173 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132842777","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Compared to traditional ring oscillator PUF (RO PUF), configurable RO PUF (CRO PUF) greatly increases the number of challenge response pairs (CRPs) and improves hardware utilization. However, in the conventional CRO PUF design, when a path is selected by the challenge to generate a response, the circuit characteristic information constituting the CRO PUF, such as the delay information of the configurable unit, the transmission model, and etc., can also be leaked. Once the adversary monitors and masters this information, they can use this information to attack the CRO PUF circuits, such as modeling attacks. This paper establishes a theoretical model of CRO PUF and analyzes its unpredictability and security. Based on this model, a new mechanism to generate the proper challenges is proposed in this paper. In the proposed mechanism, the challenge is generated and utilized by a specific way, which can delay the feature leakage of the CRO PUF, thereby improving the security of the CRO PUF.
{"title":"Theoretical Analysis of Configurable RO PUFs and Strategies to Enhance Security","authors":"Jiang Li, Hao Gao, Yijun Cui, Chenghua Wang, Yale Wang, Chongyan Gu, Weiqiang Liu","doi":"10.1109/SiPS47522.2019.9020320","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020320","url":null,"abstract":"Compared to traditional ring oscillator PUF (RO PUF), configurable RO PUF (CRO PUF) greatly increases the number of challenge response pairs (CRPs) and improves hardware utilization. However, in the conventional CRO PUF design, when a path is selected by the challenge to generate a response, the circuit characteristic information constituting the CRO PUF, such as the delay information of the configurable unit, the transmission model, and etc., can also be leaked. Once the adversary monitors and masters this information, they can use this information to attack the CRO PUF circuits, such as modeling attacks. This paper establishes a theoretical model of CRO PUF and analyzes its unpredictability and security. Based on this model, a new mechanism to generate the proper challenges is proposed in this paper. In the proposed mechanism, the challenge is generated and utilized by a specific way, which can delay the feature leakage of the CRO PUF, thereby improving the security of the CRO PUF.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"38 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129613704","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020602
Yixin Song, Yang Li, Cheng Zhang, Yongming Huang
In recent applications of millimeter wave automotive radars, the short range detection and estimation performance becomes an important design metric. Due to the sphere rather than plane form of array incoming signals, direct use of conventional spectrum or direction of arrival (DOA) estimators generally result in large performance degradation. In this paper, a naive look-up table based solution is first introduced. To solve its involved large storage requirement problem, we further transform the DOA estimation problem into the DOA classification problem, and utilize the support vector machine (SVM) framework to propose a data-driven low-complexity DOA estimator. Simulations validate the effectiveness of the propose SVM solution especially for small sample set and high storage limit.
{"title":"Data Driven Low-Complexity DOA Estimation for Ultra-Short Range Automotive Radar","authors":"Yixin Song, Yang Li, Cheng Zhang, Yongming Huang","doi":"10.1109/SiPS47522.2019.9020602","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020602","url":null,"abstract":"In recent applications of millimeter wave automotive radars, the short range detection and estimation performance becomes an important design metric. Due to the sphere rather than plane form of array incoming signals, direct use of conventional spectrum or direction of arrival (DOA) estimators generally result in large performance degradation. In this paper, a naive look-up table based solution is first introduced. To solve its involved large storage requirement problem, we further transform the DOA estimation problem into the DOA classification problem, and utilize the support vector machine (SVM) framework to propose a data-driven low-complexity DOA estimator. Simulations validate the effectiveness of the propose SVM solution especially for small sample set and high storage limit.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"42 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127090309","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020572
Chunpeng Wu, Ang Li, Bing Li, Yiran Chen
Human drivers learn driving skills from both regular (non-accidental) and accidental driving experiences, while most of current self-driving research focuses on regular driving only. We argue that learning from accidental driving data is necessary for robustly modeling driving behavior. A main challenge, however, is how accident data can be effectively used together with regular data to learn vehicle motion, since manually labeling accident data without expertise is significantly difficult. Our proposed solution for robust vehicle motion learning, in this paper, is to integrate layer-level discriminability and neuron coverage(neuron-level robustness) regulariziers into an unsupervised generative network for video prediction. Layer-level discriminability increases divergence of feature distribution between the regular data and accident data at network layers. Neuron coverage regulariziers enlarge interval span of neuron activation adopted by training samples, to reduce probability that a sample falls into untested interval regions. To accelerate training process, we propose adaptive filter reuse based on neuron coverage. Our strategies of filter reuse reduce structural network parameters, encourage memory reuse, and guarantee effectiveness of robust vehicle motion learning. Experiments results show that our model improves the inference accuracy by 1.1% compared to FCMLSTM, and cut down 10.2% training time over the traditional method with negligible accuracy loss.
{"title":"Efficiently Learning a Robust Self-Driving Model with Neuron Coverage Aware Adaptive Filter Reuse","authors":"Chunpeng Wu, Ang Li, Bing Li, Yiran Chen","doi":"10.1109/SiPS47522.2019.9020572","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020572","url":null,"abstract":"Human drivers learn driving skills from both regular (non-accidental) and accidental driving experiences, while most of current self-driving research focuses on regular driving only. We argue that learning from accidental driving data is necessary for robustly modeling driving behavior. A main challenge, however, is how accident data can be effectively used together with regular data to learn vehicle motion, since manually labeling accident data without expertise is significantly difficult. Our proposed solution for robust vehicle motion learning, in this paper, is to integrate layer-level discriminability and neuron coverage(neuron-level robustness) regulariziers into an unsupervised generative network for video prediction. Layer-level discriminability increases divergence of feature distribution between the regular data and accident data at network layers. Neuron coverage regulariziers enlarge interval span of neuron activation adopted by training samples, to reduce probability that a sample falls into untested interval regions. To accelerate training process, we propose adaptive filter reuse based on neuron coverage. Our strategies of filter reuse reduce structural network parameters, encourage memory reuse, and guarantee effectiveness of robust vehicle motion learning. Experiments results show that our model improves the inference accuracy by 1.1% compared to FCMLSTM, and cut down 10.2% training time over the traditional method with negligible accuracy loss.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"11 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127840953","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Polar code, as the first code that can probably achieve the capacity of B-DMCs, has received great attention. Belief propagation (BP) decoding algorithm, a paralleled de-coding approach for polar codes, suffers from high hardware complexity. In this paper, we devoted ourselves to proposing a data structure-based approximate BP decoder for polar code. Multiple simulations have been done. The simulation results show that by reforming the data structure of the received channel message and introducing the approximate computing schemes, significant hardware reduction has been made compared to its conventional counterpart. The hardware architecture and corresponding implementation results are also given in this paper.
{"title":"A Data Structure-Based Approximate Belief Propagation Decoder for Polar Codes","authors":"Menghui Xu, Weikang Qian, Zaichen Zhang, X. You, Chuan Zhang","doi":"10.1109/SiPS47522.2019.9020391","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020391","url":null,"abstract":"Polar code, as the first code that can probably achieve the capacity of B-DMCs, has received great attention. Belief propagation (BP) decoding algorithm, a paralleled de-coding approach for polar codes, suffers from high hardware complexity. In this paper, we devoted ourselves to proposing a data structure-based approximate BP decoder for polar code. Multiple simulations have been done. The simulation results show that by reforming the data structure of the received channel message and introducing the approximate computing schemes, significant hardware reduction has been made compared to its conventional counterpart. The hardware architecture and corresponding implementation results are also given in this paper.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"575-578 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126630419","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020469
E. C. Marques, N. Maciel, L. Naviner, Hao Cai, Jun Yang
Compressive sensing requires fewer measurements than Nyquist rate to recover sparse signals, leading to processing and energy saving. The efficiency of this technique strongly depends on the quality of the considered sparse recovery algorithm. This work focuses on a learned iterative shrinkage-thresholding algorithm where iterations are related to layers of a neural network. We analyze the performance of this algorithm for different shrinkage functions. A decrease up to 9dB in the NMSE value is achieved by choosing appropriate shrinkage function. Moreover, the estimation performance can be close to the theoretical performance bound, showing deep learning as a promising tool for sparse signal estimation. This work can be applied in several areas such as image processing, Internet of Things (IoT), cognitive radio networks, and sparse channel estimation for wireless communications.
{"title":"Nonlinear Functions in Learned Iterative Shrinkage-Thresholding Algorithm for Sparse Signal Recovery","authors":"E. C. Marques, N. Maciel, L. Naviner, Hao Cai, Jun Yang","doi":"10.1109/SiPS47522.2019.9020469","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020469","url":null,"abstract":"Compressive sensing requires fewer measurements than Nyquist rate to recover sparse signals, leading to processing and energy saving. The efficiency of this technique strongly depends on the quality of the considered sparse recovery algorithm. This work focuses on a learned iterative shrinkage-thresholding algorithm where iterations are related to layers of a neural network. We analyze the performance of this algorithm for different shrinkage functions. A decrease up to 9dB in the NMSE value is achieved by choosing appropriate shrinkage function. Moreover, the estimation performance can be close to the theoretical performance bound, showing deep learning as a promising tool for sparse signal estimation. This work can be applied in several areas such as image processing, Internet of Things (IoT), cognitive radio networks, and sparse channel estimation for wireless communications.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"66 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117345528","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020535
P. Shvartsman, Xinmiao Zhang
The Advanced Encryption Standard (AES) is the current standard for symmetric key cipher and is algorithmically secure. Side channel attacks that target power consumption can reveal the secret key in AES implementations. Masking data with random variables is one of the main methods used to thwart power analysis attacks. Data can be masked with multiple random variables to prevent higher-order attacks at the cost of a large increase in area. A novel masking scheme for AES resistant to second-order attacks is proposed. Instead of using an additional mask, variation in finite field construction is exploited to increase resistance to second-order attacks. As a result, the area requirement is reduced. For an example AES encryptor, the proposed design is 12% smaller compared to the previous best design, with a very small drop in achievable security level.
{"title":"Side Channel Attack Resistant AES Design Based on Finite Field Construction Variation","authors":"P. Shvartsman, Xinmiao Zhang","doi":"10.1109/SiPS47522.2019.9020535","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020535","url":null,"abstract":"The Advanced Encryption Standard (AES) is the current standard for symmetric key cipher and is algorithmically secure. Side channel attacks that target power consumption can reveal the secret key in AES implementations. Masking data with random variables is one of the main methods used to thwart power analysis attacks. Data can be masked with multiple random variables to prevent higher-order attacks at the cost of a large increase in area. A novel masking scheme for AES resistant to second-order attacks is proposed. Instead of using an additional mask, variation in finite field construction is exploited to increase resistance to second-order attacks. As a result, the area requirement is reduced. For an example AES encryptor, the proposed design is 12% smaller compared to the previous best design, with a very small drop in achievable security level.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124070055","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020579
Emad A. Ibrahim, Min Li, J. P. D. Gyvez
Targeting ultrasound-based gesture recognition, this paper proposes a new universal PRESS/HOLD/RELEASE approach that leverages the diversity of gestures performed on smart devices such as mobile phones and IoT nodes. The new set of gestures are generated by interleaving PRESS/HOLD/RELEASE patterns; abbreviated as P/H/R, with gestures like sweeps between a number of microphones. P/H/R patterns are constructed by a hand as it approaches a top of a microphone to generate a virtual Press. After that, the hand settles for an undefined period of time to generate a virtual Hold and finally departs to generate a virtual Release. The same hand can sweep to a 2nd microphone and perform another P/H/R. Interleaving the P/H/R patterns expands the number of performed gestures. Assuming an on-board speaker transmitting ultrasonic signals, the detection is performed on Doppler shift readings generated by a hand as it approaches and departs a top of a microphone. The Doppler shift readings are presented in a sequence of down-mixed ultrasonic spectrogram frames. We train a Temporal Convolutional Network (TCN) to classify the P/H/R patterns under different environmental noises. Our experimental results show that such P/H/R patterns at a top of a microphone can be achieved with 96.6% accuracy under different noise conditions. A group of P/H/R based gestures has been tested on commercially off-the-shelf (COTS) Samsung Galaxy S7 Edge. Different P/H/R interleaved gestures (such as sweeps, long taps, etc.) are designed using two microphones and a single speaker while using as low as $sim 5mathrm{K}$ parameters and as low as $sim 0.15$ Million operations (MOPs) in compute power per inference. The P/H/R interleaved set of gestures are intuitive and hence are easy to learn by end users. This paves its way to be deployed by smartphones and smart speakers for mass production.
{"title":"PRESS/HOLD/RELEASE Ultrasonic Gestures and Low Complexity Recognition Based on TCN","authors":"Emad A. Ibrahim, Min Li, J. P. D. Gyvez","doi":"10.1109/SiPS47522.2019.9020579","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020579","url":null,"abstract":"Targeting ultrasound-based gesture recognition, this paper proposes a new universal PRESS/HOLD/RELEASE approach that leverages the diversity of gestures performed on smart devices such as mobile phones and IoT nodes. The new set of gestures are generated by interleaving PRESS/HOLD/RELEASE patterns; abbreviated as P/H/R, with gestures like sweeps between a number of microphones. P/H/R patterns are constructed by a hand as it approaches a top of a microphone to generate a virtual Press. After that, the hand settles for an undefined period of time to generate a virtual Hold and finally departs to generate a virtual Release. The same hand can sweep to a 2nd microphone and perform another P/H/R. Interleaving the P/H/R patterns expands the number of performed gestures. Assuming an on-board speaker transmitting ultrasonic signals, the detection is performed on Doppler shift readings generated by a hand as it approaches and departs a top of a microphone. The Doppler shift readings are presented in a sequence of down-mixed ultrasonic spectrogram frames. We train a Temporal Convolutional Network (TCN) to classify the P/H/R patterns under different environmental noises. Our experimental results show that such P/H/R patterns at a top of a microphone can be achieved with 96.6% accuracy under different noise conditions. A group of P/H/R based gestures has been tested on commercially off-the-shelf (COTS) Samsung Galaxy S7 Edge. Different P/H/R interleaved gestures (such as sweeps, long taps, etc.) are designed using two microphones and a single speaker while using as low as $sim 5mathrm{K}$ parameters and as low as $sim 0.15$ Million operations (MOPs) in compute power per inference. The P/H/R interleaved set of gestures are intuitive and hence are easy to learn by end users. This paves its way to be deployed by smartphones and smart speakers for mass production.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"53 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122717021","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020464
Saman Payvar, Mir Khan, Rafael Stahl, Daniel Mueller-Gritschneder, J. Boutellier
Convolutional Neural Networks (CNNs) have previously provided unforeseen results in automatic image analysis and interpretation, an area which has numerous applications in both consumer electronics and industry. However, the signal processing related to CNNs is computationally very demanding, which has prohibited their use in the smallest embedded computing platforms, to which many Internet of Things (IoT) devices belong. Fortunately, in the recent years researchers have developed many approaches for optimizing the performance and for shrinking the memory footprint of CNNs. This paper presents a neural-network-based image classifier that has been trained to classify vehicle images into four different classes. The neural network is optimized by a technique called binarization, and the resulting binarized network is placed to an IoT-class processor core for execution. Binarization reduces the memory footprint of the CNN by around 95% and increases performance by more than $6 times $. Furthermore, we show that by utilizing a custom instruction ’popcount’ of the processor, the performance of the binarized vehicle classifier can still be increased by more than $2 times $, making the CNN-based image classifier suitable for the smallest embedded processors.
{"title":"Neural Network-based Vehicle Image Classification for IoT Devices","authors":"Saman Payvar, Mir Khan, Rafael Stahl, Daniel Mueller-Gritschneder, J. Boutellier","doi":"10.1109/SiPS47522.2019.9020464","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020464","url":null,"abstract":"Convolutional Neural Networks (CNNs) have previously provided unforeseen results in automatic image analysis and interpretation, an area which has numerous applications in both consumer electronics and industry. However, the signal processing related to CNNs is computationally very demanding, which has prohibited their use in the smallest embedded computing platforms, to which many Internet of Things (IoT) devices belong. Fortunately, in the recent years researchers have developed many approaches for optimizing the performance and for shrinking the memory footprint of CNNs. This paper presents a neural-network-based image classifier that has been trained to classify vehicle images into four different classes. The neural network is optimized by a technique called binarization, and the resulting binarized network is placed to an IoT-class processor core for execution. Binarization reduces the memory footprint of the CNN by around 95% and increases performance by more than $6 times $. Furthermore, we show that by utilizing a custom instruction ’popcount’ of the processor, the performance of the binarized vehicle classifier can still be increased by more than $2 times $, making the CNN-based image classifier suitable for the smallest embedded processors.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"7 5 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126140675","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2019-10-01DOI: 10.1109/SiPS47522.2019.9020321
Cheng-Ping Hsieh, Yi-Ta Chen, Win-Ken Beh, A. Wu
Since stress has a strong influence on human’s health, it is necessary to automatically detect stress in our daily life. In this paper, we aim to improve the performance and obtain the dominant features in stress detection based on Electrodermal Activity (EDA). Compared to the methods in Wearable Stress and Affect Dataset (WESAD), we propose several enhancements to get higher f1-scores, including less overlapped signal segmentation, more signal processing features, and extreme gradient boosting classification algorithm (XGBoost). Furthermore, we select dominant features according to their importance in classifier and correlation among other features while keeping high performance. Experiment results show that with 9 dominant features in XGBoost, we can achieve 92.38% (+ 17.87%) and 89.92% (+14.58%) f1-scores compared to WESAD on chest-and wrist-based EDA signal respectively. The features we choose suggest that the magnitude of low frequency and the complexity of high frequency EDA signal contain the most significant information in stress detection.
{"title":"Feature Selection Framework for XGBoost Based on Electrodermal Activity in Stress Detection","authors":"Cheng-Ping Hsieh, Yi-Ta Chen, Win-Ken Beh, A. Wu","doi":"10.1109/SiPS47522.2019.9020321","DOIUrl":"https://doi.org/10.1109/SiPS47522.2019.9020321","url":null,"abstract":"Since stress has a strong influence on human’s health, it is necessary to automatically detect stress in our daily life. In this paper, we aim to improve the performance and obtain the dominant features in stress detection based on Electrodermal Activity (EDA). Compared to the methods in Wearable Stress and Affect Dataset (WESAD), we propose several enhancements to get higher f1-scores, including less overlapped signal segmentation, more signal processing features, and extreme gradient boosting classification algorithm (XGBoost). Furthermore, we select dominant features according to their importance in classifier and correlation among other features while keeping high performance. Experiment results show that with 9 dominant features in XGBoost, we can achieve 92.38% (+ 17.87%) and 89.92% (+14.58%) f1-scores compared to WESAD on chest-and wrist-based EDA signal respectively. The features we choose suggest that the magnitude of low frequency and the complexity of high frequency EDA signal contain the most significant information in stress detection.","PeriodicalId":256971,"journal":{"name":"2019 IEEE International Workshop on Signal Processing Systems (SiPS)","volume":"22 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125906486","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：