在流式识别方法中,分块识别破坏并行性且消耗资源较大,而限制自注意力机制的上下文识别很难获得所有信息.由此,文中提出轻量化端到端声学架构(CFLASH-Transducer).为了获取细腻的局部特征,采用轻量化的FLASH(Fast Linear Attention with...在流式识别方法中,分块识别破坏并行性且消耗资源较大,而限制自注意力机制的上下文识别很难获得所有信息.由此,文中提出轻量化端到端声学架构(CFLASH-Transducer).为了获取细腻的局部特征,采用轻量化的FLASH(Fast Linear Attention with a Single Head)与卷积神经网络块结合.卷积块中采用Inception V2网络,提取语音信号多尺度的局部特征.再通过Coordinate Attention机制捕获特征的位置信息和多通道之间的相互关联.此外,采用深度可分离卷积,用于特征增强和层间平滑过渡.为了使其可流式化处理音频,采用RNN-T(Recurrent Neural Network Transducer)架构进行训练与解码.将当前块已经计算的全局注意力作为隐变量,传入后续块中,串联各块信息,保留训练的并行性和相关性,并且不会随着序列的增长而消耗计算资源.在开源数据集THCHS30上进行训练与测试,CFLASH-Transducer取得较高的识别率.并且相比离线识别,流式识别精度损失不超过1%.展开更多
Upset errors in 90-nm 64 Mb NOR-type floating-gate Flash memory induced by accelerated ^（129）Xe and ^（209）Bi ions are investigated in detail. The linear energy transfer covers the range from 50 to 99.8 Me V/（mg/c...Upset errors in 90-nm 64 Mb NOR-type floating-gate Flash memory induced by accelerated ^（129）Xe and ^（209）Bi ions are investigated in detail. The linear energy transfer covers the range from 50 to 99.8 Me V/（mg/cm^2）. When the memory chips are powered off during heavy ions irradiation, single-event-latch-up and single-event-function-interruption are excluded,and only 0-〉1 upset errors in the memory array are observed. These error bit rates seem very difficult to achieve and cannot be simply recovered based on the power cycle. The number of error bits shows a strong dependence on the linear energy transfer（LET）. Under room-temperature annealing conditions, the upset errors can be reduced by about two orders of magnitude using rewrite/reprogram operations, but they subsequently increase once again in a few minutes after the power cycle. High-temperature annealing can diminish almost all error bits, which are affected by the lower LET ^（129）Xe ions. The percolation path between the floating-gate（FG） and the substrate contributes to the radiation-induced leakage current, and has been identified as the root cause of the upset errors of the Flash memory array in this work.展开更多
文摘在流式识别方法中,分块识别破坏并行性且消耗资源较大,而限制自注意力机制的上下文识别很难获得所有信息.由此,文中提出轻量化端到端声学架构(CFLASH-Transducer).为了获取细腻的局部特征,采用轻量化的FLASH(Fast Linear Attention with a Single Head)与卷积神经网络块结合.卷积块中采用Inception V2网络,提取语音信号多尺度的局部特征.再通过Coordinate Attention机制捕获特征的位置信息和多通道之间的相互关联.此外,采用深度可分离卷积,用于特征增强和层间平滑过渡.为了使其可流式化处理音频,采用RNN-T(Recurrent Neural Network Transducer)架构进行训练与解码.将当前块已经计算的全局注意力作为隐变量,传入后续块中,串联各块信息,保留训练的并行性和相关性,并且不会随着序列的增长而消耗计算资源.在开源数据集THCHS30上进行训练与测试,CFLASH-Transducer取得较高的识别率.并且相比离线识别,流式识别精度损失不超过1%.
基金Project supported by the National Natural Science Foundation of China(Grant No.616340084)the Youth Innovation Promotion Association of CAS(Grant No.2014101)+1 种基金the International Cooperation Project of CASthe Austrian-Chinese Cooperative R&D Projects(Grant No.172511KYSB20150006)
文摘Upset errors in 90-nm 64 Mb NOR-type floating-gate Flash memory induced by accelerated ^（129）Xe and ^（209）Bi ions are investigated in detail. The linear energy transfer covers the range from 50 to 99.8 Me V/（mg/cm^2）. When the memory chips are powered off during heavy ions irradiation, single-event-latch-up and single-event-function-interruption are excluded,and only 0-〉1 upset errors in the memory array are observed. These error bit rates seem very difficult to achieve and cannot be simply recovered based on the power cycle. The number of error bits shows a strong dependence on the linear energy transfer（LET）. Under room-temperature annealing conditions, the upset errors can be reduced by about two orders of magnitude using rewrite/reprogram operations, but they subsequently increase once again in a few minutes after the power cycle. High-temperature annealing can diminish almost all error bits, which are affected by the lower LET ^（129）Xe ions. The percolation path between the floating-gate（FG） and the substrate contributes to the radiation-induced leakage current, and has been identified as the root cause of the upset errors of the Flash memory array in this work.
