In this paper, a novel multi-scroll memristive chaotic system is designed based on Chua's system via introducing a nonlinear memristor. The dynamics of this system is analyzed based on bifurcation diagrams, Lyapunov exponents and phase diagrams. Subsequently, an image encryption scheme based on this system is then proposed. First, the proposed chaotic system is used to generate continuously robust chaotic sequences, the hash values of plaintext images are embedded in the generation and selection of chaotic sequences and involved in each step of encryption to establish the coupling relationship between plaintext and ciphertext. Second, Knuth-Durstenfeld algorithm is used to scramble the high four-bit plane of the plain image twice, and the chaotic sequence is used as the index sequence, which greatly improves the efficiency and randomness of the permutation process. Finally, chaotic sequences are involved in DNA coding rules and pixel-level diffusion. The algorithm is highly sensitive to plain images, and it can realize adaptive encryption. Through performance analysis and comparison with recent literature, the proposed algorithm can cope with various attacks and show excellent performance.
{"title":"An Encryption Scheme Using Multi-Scroll Memristive Chaotic System","authors":"Fan Wu, Musha Ji E, Lidan Wang, Shukai Duan","doi":"10.1145/3590003.3590105","DOIUrl":"https://doi.org/10.1145/3590003.3590105","url":null,"abstract":"In this paper, a novel multi-scroll memristive chaotic system is designed based on Chua's system via introducing a nonlinear memristor. The dynamics of this system is analyzed based on bifurcation diagrams, Lyapunov exponents and phase diagrams. Subsequently, an image encryption scheme based on this system is then proposed. First, the proposed chaotic system is used to generate continuously robust chaotic sequences, the hash values of plaintext images are embedded in the generation and selection of chaotic sequences and involved in each step of encryption to establish the coupling relationship between plaintext and ciphertext. Second, Knuth-Durstenfeld algorithm is used to scramble the high four-bit plane of the plain image twice, and the chaotic sequence is used as the index sequence, which greatly improves the efficiency and randomness of the permutation process. Finally, chaotic sequences are involved in DNA coding rules and pixel-level diffusion. The algorithm is highly sensitive to plain images, and it can realize adaptive encryption. Through performance analysis and comparison with recent literature, the proposed algorithm can cope with various attacks and show excellent performance.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"16 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115440888","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}
Abstract: Virtual humans created by computers using deep learning technology are being used widely in a variety of fields, including personal assistance, intelligent customer service, and online education. Human-computer interaction systems integrate multi-modal technologies like speech recognition, dialogue systems, speech synthesis, and virtual digital human video synthesis as one of the applications of virtual humans. In this paper, we first design the framework for a human-computer interaction system based on a virtual human; next, we classify the talking head video synthesis model according to the generation of a virtual human's depth; finally, we conduct a systematic review of the technical developments in talking head video generation over the last five years, highlighting seminal work.
{"title":"Virtual Human Talking-Head Generation","authors":"Wenchao Song, Qiang He, Guowei Chen","doi":"10.1145/3590003.3590004","DOIUrl":"https://doi.org/10.1145/3590003.3590004","url":null,"abstract":"Abstract: Virtual humans created by computers using deep learning technology are being used widely in a variety of fields, including personal assistance, intelligent customer service, and online education. Human-computer interaction systems integrate multi-modal technologies like speech recognition, dialogue systems, speech synthesis, and virtual digital human video synthesis as one of the applications of virtual humans. In this paper, we first design the framework for a human-computer interaction system based on a virtual human; next, we classify the talking head video synthesis model according to the generation of a virtual human's depth; finally, we conduct a systematic review of the technical developments in talking head video generation over the last five years, highlighting seminal work.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128083968","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}
Aiming at the problems of storage, batch migration and centralized processing of visual digital images of infrared imaging products, this paper takes digital image noise reduction as the main research object and starts with the concept of image partial differential equation processing. Based on the development history, advantages, practicability and operability of digital image processing by partial differential equation, it is concluded that digital image processing technology based on P-M model method is more suitable for modern image processing, and also broadens and improves the basic algorithm of digital image processing in the past. On this basis, the image quality is evaluated by using the fuzzy comprehensive evaluation theory based on analytic hierarchy process. The results show that the optimized processing system can screen the advantages and disadvantages of visual digital images of infrared imaging products and provide technical support.
{"title":"Digital image denoising by partial differential equation based on P-M model and its fuzzy evaluation method system","authors":"Jingying Liu, Yang Hu","doi":"10.1145/3590003.3590091","DOIUrl":"https://doi.org/10.1145/3590003.3590091","url":null,"abstract":"Aiming at the problems of storage, batch migration and centralized processing of visual digital images of infrared imaging products, this paper takes digital image noise reduction as the main research object and starts with the concept of image partial differential equation processing. Based on the development history, advantages, practicability and operability of digital image processing by partial differential equation, it is concluded that digital image processing technology based on P-M model method is more suitable for modern image processing, and also broadens and improves the basic algorithm of digital image processing in the past. On this basis, the image quality is evaluated by using the fuzzy comprehensive evaluation theory based on analytic hierarchy process. The results show that the optimized processing system can screen the advantages and disadvantages of visual digital images of infrared imaging products and provide technical support.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"26 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123784544","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}
The Discrete Hopfield Neural Network introduces a G-Type Random 3 Satisfiability logic structure, which can improve the flexibility of the logic structure and meet the requirements of all combinatorial problems. Usually, Exhaustive Search (ES) is regarded as the basic learning algorithm to search the fitness of neurons. To improve the efficiency of the learning algorithm. In this paper, we introduce the Estimation of Distribution Algorithm (EDA) as a learning algorithm for the model. To study the learning mechanism of EDA to improve search efficiency, this study focuses on the impact of EDA on the model under different proportions of literals and evaluates the performance of the model at different phases through evaluation indicators. Analyze the effect of EDA on the synaptic weights and the global solution. From the discussion, it can be found that compared with ES, EDA has a larger search space at the same efficiency, which makes the probability of obtaining satisfactory weights higher, and the proportion of global solutions obtained is higher. Higher proportions of positive literals help to improve the model performance.
{"title":"Estimation of Distribution Algorithm with Discrete Hopfield Neural Network for GRAN3SAT Analysis","authors":"Yuan Gao, Chengfeng Zheng, Ju Chen, Yueling Guo","doi":"10.1145/3590003.3590021","DOIUrl":"https://doi.org/10.1145/3590003.3590021","url":null,"abstract":"The Discrete Hopfield Neural Network introduces a G-Type Random 3 Satisfiability logic structure, which can improve the flexibility of the logic structure and meet the requirements of all combinatorial problems. Usually, Exhaustive Search (ES) is regarded as the basic learning algorithm to search the fitness of neurons. To improve the efficiency of the learning algorithm. In this paper, we introduce the Estimation of Distribution Algorithm (EDA) as a learning algorithm for the model. To study the learning mechanism of EDA to improve search efficiency, this study focuses on the impact of EDA on the model under different proportions of literals and evaluates the performance of the model at different phases through evaluation indicators. Analyze the effect of EDA on the synaptic weights and the global solution. From the discussion, it can be found that compared with ES, EDA has a larger search space at the same efficiency, which makes the probability of obtaining satisfactory weights higher, and the proportion of global solutions obtained is higher. Higher proportions of positive literals help to improve the model performance.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129780673","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}
Yongcai Tao, Jitao Zhang, Lin Wei, Yufei Gao, Lei Shi
Abstract—At present, machine learning and deep learning are often used for network traffic intrusion detection. In order to solve the problem of unfocused feature extraction in these methods and improve the accuracy of network intrusion detection, this paper proposes an intrusion detection model that combines Attention and BiLSTM-DNN(ABD). The model uses Attention to perform preliminary feature extraction on input data, reads the relationship between different features, then uses BiLSTM to extract long-distance dependent features, uses DNN to further extract deep-level features, and finally obtains classification through SoftMax classifier. The comparison experiment uses the NSL_KDD data set, and models such as BiLSTM-DNN, support vector machine, decision tree and random forest are selected as the comparison experiment model. The experimental results show that the accuracy of the ABD is improved by 1.0% and 2.0% on the two-category and five-category tasks, respectively, which verifies the effectiveness of the method.
{"title":"An Intrusion Detection Model With Attention and BiLSTM-DNN","authors":"Yongcai Tao, Jitao Zhang, Lin Wei, Yufei Gao, Lei Shi","doi":"10.1145/3590003.3590018","DOIUrl":"https://doi.org/10.1145/3590003.3590018","url":null,"abstract":"Abstract—At present, machine learning and deep learning are often used for network traffic intrusion detection. In order to solve the problem of unfocused feature extraction in these methods and improve the accuracy of network intrusion detection, this paper proposes an intrusion detection model that combines Attention and BiLSTM-DNN(ABD). The model uses Attention to perform preliminary feature extraction on input data, reads the relationship between different features, then uses BiLSTM to extract long-distance dependent features, uses DNN to further extract deep-level features, and finally obtains classification through SoftMax classifier. The comparison experiment uses the NSL_KDD data set, and models such as BiLSTM-DNN, support vector machine, decision tree and random forest are selected as the comparison experiment model. The experimental results show that the accuracy of the ABD is improved by 1.0% and 2.0% on the two-category and five-category tasks, respectively, which verifies the effectiveness of the method.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"172 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130755770","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}
Animal pollinators have been supporting the lives of human beings on Earth. Bee pollinators are the biggest contributors to the pollination of crops, providing humans with food. Given such circumstances, this paper investigates the population of honeybee colonies and the processes of bee pollination. We constructed Honeybee Colony Population Model (BCPM) to predict the population of a honeybee colony over time. We first outlined the life cycle of a honeybee, including eggs, larval stage, pupal stage, and adult bee stage. Within the adult bee stage, bees transition back and forth between foragers and hive bees depending on the number of available resources and the workload of nursing tasks. By listing out factors that affect the population in each stage, we established equations representing the rate of change in each of the stages of a honeybee's life cycle, as well as an equation describing the change in resource storage. We also evaluated the death rate and the resources in each month of the year and calculated each group's typical maximum, minimum, and mean population in a honeybee colony: 3 years after the establishment of the colony, the total adult population follows a seasonal change with recurring patterns each year, giving a maximum of 100862 bees and a minimum of 35676 bees. The annual average population is found to be 64877 bees. We then conducted a sensitivity analysis on BCPM and found that the initial number of bee hives and the initial amount of available resources have the most significant impact on the population of the colony. We also observed an unusual pattern in the cross-analysis of the two factors and constructed Simplified Colony Collapse Disorder Model (SCCDM) to predict whether a colony will collapse using only one equation. In response to estimate the number of hives needed to support the pollination of a specific land area, we constructed Hive Deployment Model (HDM). We first divided the land into 20 nodes and then found the most appropriate locations to place the hives. After establishing the equations for movements between nodes per day per forager group, we developed an iterating algorithm to find the number of hives needed to pollinate crops on 20 acres of land. We collected data for 9 typical bee-pollinated plants and found the number of hives needed for each type of plant based on the algorithm, with blueberries being the most demanding, requiring 83 hives, whilst apples and roses only required 2 hives at the other end of the spectrum. Then, we established a sensitivity analysis to ensure the stability of the model by changing two arbitrary parameters. Finally, we discussed the potential advantages and disadvantages of our model. We have also created a non-technical blog that summarizes our investigation, presenting our results in a simplified way
{"title":"Mathematical models of colony population dynamics and hive placement","authors":"Zixuan Zhang, Dongyi He, Hanwen Zhang","doi":"10.1145/3590003.3590070","DOIUrl":"https://doi.org/10.1145/3590003.3590070","url":null,"abstract":"Animal pollinators have been supporting the lives of human beings on Earth. Bee pollinators are the biggest contributors to the pollination of crops, providing humans with food. Given such circumstances, this paper investigates the population of honeybee colonies and the processes of bee pollination. We constructed Honeybee Colony Population Model (BCPM) to predict the population of a honeybee colony over time. We first outlined the life cycle of a honeybee, including eggs, larval stage, pupal stage, and adult bee stage. Within the adult bee stage, bees transition back and forth between foragers and hive bees depending on the number of available resources and the workload of nursing tasks. By listing out factors that affect the population in each stage, we established equations representing the rate of change in each of the stages of a honeybee's life cycle, as well as an equation describing the change in resource storage. We also evaluated the death rate and the resources in each month of the year and calculated each group's typical maximum, minimum, and mean population in a honeybee colony: 3 years after the establishment of the colony, the total adult population follows a seasonal change with recurring patterns each year, giving a maximum of 100862 bees and a minimum of 35676 bees. The annual average population is found to be 64877 bees. We then conducted a sensitivity analysis on BCPM and found that the initial number of bee hives and the initial amount of available resources have the most significant impact on the population of the colony. We also observed an unusual pattern in the cross-analysis of the two factors and constructed Simplified Colony Collapse Disorder Model (SCCDM) to predict whether a colony will collapse using only one equation. In response to estimate the number of hives needed to support the pollination of a specific land area, we constructed Hive Deployment Model (HDM). We first divided the land into 20 nodes and then found the most appropriate locations to place the hives. After establishing the equations for movements between nodes per day per forager group, we developed an iterating algorithm to find the number of hives needed to pollinate crops on 20 acres of land. We collected data for 9 typical bee-pollinated plants and found the number of hives needed for each type of plant based on the algorithm, with blueberries being the most demanding, requiring 83 hives, whilst apples and roses only required 2 hives at the other end of the spectrum. Then, we established a sensitivity analysis to ensure the stability of the model by changing two arbitrary parameters. Finally, we discussed the potential advantages and disadvantages of our model. We have also created a non-technical blog that summarizes our investigation, presenting our results in a simplified way","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130834943","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}
Abstract: The filtered feature selection method has low computational complexity and less time, and is widely used in feature selection, but the filtered method only considers the importance of features for label classification and ignores the correlation between features. For this reason, a feature selection method with improved principal component analysis is proposed. The main idea of the method is that on the basis of principal components, the loadings of each indicator on different principal components and their variance contribution ratios with that principal component are considered. A number of indicators with the largest cumulative contribution rates were selected, so that the final extracted indicators retained more information. Subsequently, comparative experiments are conducted using the UCI dataset, and the results show that the approach proposed in this paper has some superiority over other methods. Finally, the features of China's green innovation efficiency are selected using the approach proposed in this paper to demonstrate the feasibility of the method.
{"title":"Feature selection based on improved principal component analysis","authors":"Zhang Li, Yihui Qiu","doi":"10.1145/3590003.3590036","DOIUrl":"https://doi.org/10.1145/3590003.3590036","url":null,"abstract":"Abstract: The filtered feature selection method has low computational complexity and less time, and is widely used in feature selection, but the filtered method only considers the importance of features for label classification and ignores the correlation between features. For this reason, a feature selection method with improved principal component analysis is proposed. The main idea of the method is that on the basis of principal components, the loadings of each indicator on different principal components and their variance contribution ratios with that principal component are considered. A number of indicators with the largest cumulative contribution rates were selected, so that the final extracted indicators retained more information. Subsequently, comparative experiments are conducted using the UCI dataset, and the results show that the approach proposed in this paper has some superiority over other methods. Finally, the features of China's green innovation efficiency are selected using the approach proposed in this paper to demonstrate the feasibility of the method.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"232 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131118754","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}
Yaozheng Xing, Jie Yuan, Qixun Liu, Shihao Peng, Yan Yan, Junyi Yao
To address the problems of many parameters and loss of spatial information in traditional Unet networks, this paper proposes a U-Net-based brain tumor segmentation model named MM-UNet to solve the problem of 3D image segmentation. Firstly, the U-Net model performs three times downsampling to extract the image features for the changing characteristics of brain tumor 3D images, which reduces the number of model parameters while maximally preserving the target edge features; then, a structure similar to FPN was used to achieve the fusion of multi-scale predictions; we introduce the channel attention mechanism and pixel attention mechanism to establish the relationship between global features; meanwhile, to improve the generalization ability of the model, data augmentation techniques are used to enhance the information. The experimental results show that the model proposed in this paper has improved the accuracy of brain tumor segmentation compared with U- Net, PSPNet, ICNet, and Fast- SCNN, suggesting 3.9%, 1.3%, 5%, and 3.9%, respectively.
{"title":"MM-UNet: Multi-attention mechanism and multi-scale feature fusion UNet for tumor image segmentation","authors":"Yaozheng Xing, Jie Yuan, Qixun Liu, Shihao Peng, Yan Yan, Junyi Yao","doi":"10.1145/3590003.3590047","DOIUrl":"https://doi.org/10.1145/3590003.3590047","url":null,"abstract":"To address the problems of many parameters and loss of spatial information in traditional Unet networks, this paper proposes a U-Net-based brain tumor segmentation model named MM-UNet to solve the problem of 3D image segmentation. Firstly, the U-Net model performs three times downsampling to extract the image features for the changing characteristics of brain tumor 3D images, which reduces the number of model parameters while maximally preserving the target edge features; then, a structure similar to FPN was used to achieve the fusion of multi-scale predictions; we introduce the channel attention mechanism and pixel attention mechanism to establish the relationship between global features; meanwhile, to improve the generalization ability of the model, data augmentation techniques are used to enhance the information. The experimental results show that the model proposed in this paper has improved the accuracy of brain tumor segmentation compared with U- Net, PSPNet, ICNet, and Fast- SCNN, suggesting 3.9%, 1.3%, 5%, and 3.9%, respectively.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129477044","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}
Che Wang, Jifeng Hu, Fuhu Song, Jiao Huang, Zixuan Yang, Yusen Wang
Deep reinforcement learning has achieved significant success in solving sequential decision-making tasks. Excellent models usually require the input of valid state signals during training, which is challenging to encode temporal state features for the deep reinforcement learning model. To address this issue, recent methods attempt to encode multi-step sequential state signals so as to obtain more comprehensive observational information. However, these methods usually have a lower performance on complex continuous control tasks because mapping the state sequence into a low-dimensional embedding causes blurring of the immediate state features. In this paper, we propose a multiple frequency bands temporal state representation learning framework. The temporal state signals are decomposed into discrete state signals of various frequency bands by Discrete Fourier Transform (DFT). Then, feature signals filtered out different high-frequency bands are generated. Meanwhile, the mask generator evaluates the weights of signals of various frequency bands and encodes high-quality representations for agent training. Our intuition is that temporal state representations considering multiple frequency bands have high fidelity and stability. We conduct experiments tasks and verify that our method has obvious advantages over the baseline in complex continuous control tasks such as Walker and Crawler.
{"title":"Multiple Frequency Bands Temporal State Representation for Deep Reinforcement Learning","authors":"Che Wang, Jifeng Hu, Fuhu Song, Jiao Huang, Zixuan Yang, Yusen Wang","doi":"10.1145/3590003.3590058","DOIUrl":"https://doi.org/10.1145/3590003.3590058","url":null,"abstract":"Deep reinforcement learning has achieved significant success in solving sequential decision-making tasks. Excellent models usually require the input of valid state signals during training, which is challenging to encode temporal state features for the deep reinforcement learning model. To address this issue, recent methods attempt to encode multi-step sequential state signals so as to obtain more comprehensive observational information. However, these methods usually have a lower performance on complex continuous control tasks because mapping the state sequence into a low-dimensional embedding causes blurring of the immediate state features. In this paper, we propose a multiple frequency bands temporal state representation learning framework. The temporal state signals are decomposed into discrete state signals of various frequency bands by Discrete Fourier Transform (DFT). Then, feature signals filtered out different high-frequency bands are generated. Meanwhile, the mask generator evaluates the weights of signals of various frequency bands and encodes high-quality representations for agent training. Our intuition is that temporal state representations considering multiple frequency bands have high fidelity and stability. We conduct experiments tasks and verify that our method has obvious advantages over the baseline in complex continuous control tasks such as Walker and Crawler.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124679109","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}
The research and development of deep learning cannot be separated from deep neural networks (DNNs). DNNs become deeper and more complex in pursuit of accuracy and precision, leading to significantly increasing inference time and training cost. Existing deep learning frameworks optimize a DNN to improve its runtime performance by transforming computational graphs based on hand-written rules. It is hard to scale when adding some new operators into DNNs. TASO can automatically generate graph substitutions that solve maintainability problems. An optimized graph will be explored by applying a sequence of graph substitutions. However, TASO only considers the runtime performance of the model during the search, which may lose potential optimization. We propose HeuSO, a fine-grained computational graph optimizer with heuristics to handle this problem. HeuSO extracts the type and number of operators of the computational graph and classifies them into four abstract types as high-level features, which facilitate subsequent heuristic search and pruning algorithms. HeuSO generates a better sequence of graph substitutions and finds a better-optimized graph by the heuristic function, which integrates the cost and high-level features of the model. To further reduce the time of searching, HeuSO implements a pruning algorithm. Through high-level specifications, HeuSO can quickly determine whether subgraphs of the original graph match the substitution rules. Evaluations on seven DNNs demonstrate that HeuSO outperforms state-of-the-art frameworks with 2.35 × speedup while accelerating search time by up to 1.58 ×.
{"title":"Heuristic Search for DNN Graph Substitutions","authors":"Feifei Deng, Hongkang Liu","doi":"10.1145/3590003.3590044","DOIUrl":"https://doi.org/10.1145/3590003.3590044","url":null,"abstract":"The research and development of deep learning cannot be separated from deep neural networks (DNNs). DNNs become deeper and more complex in pursuit of accuracy and precision, leading to significantly increasing inference time and training cost. Existing deep learning frameworks optimize a DNN to improve its runtime performance by transforming computational graphs based on hand-written rules. It is hard to scale when adding some new operators into DNNs. TASO can automatically generate graph substitutions that solve maintainability problems. An optimized graph will be explored by applying a sequence of graph substitutions. However, TASO only considers the runtime performance of the model during the search, which may lose potential optimization. We propose HeuSO, a fine-grained computational graph optimizer with heuristics to handle this problem. HeuSO extracts the type and number of operators of the computational graph and classifies them into four abstract types as high-level features, which facilitate subsequent heuristic search and pruning algorithms. HeuSO generates a better sequence of graph substitutions and finds a better-optimized graph by the heuristic function, which integrates the cost and high-level features of the model. To further reduce the time of searching, HeuSO implements a pruning algorithm. Through high-level specifications, HeuSO can quickly determine whether subgraphs of the original graph match the substitution rules. Evaluations on seven DNNs demonstrate that HeuSO outperforms state-of-the-art frameworks with 2.35 × speedup while accelerating search time by up to 1.58 ×.","PeriodicalId":340225,"journal":{"name":"Proceedings of the 2023 2nd Asia Conference on Algorithms, Computing and Machine Learning","volume":"12 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-03-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127794502","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：