SE-MCNN-CTC的中文语音识别声学模型

针对国内外缺少对振动轮噪声预估的问题,以某型振动轮为研究对象,首先基于动力学有限元理论对振动轮进行频率响应分析,其次采用声学边界元技术对振动轮辐射噪声进行了数值模拟,并通过实验验证了仿真结果的准确性,然后比较了垂直振动与圆周振动两种不同激振形式对辐射噪声的影响,得出垂直振动辐射噪声低的结论,最后对驾驶室声腔模态进行了仿真,与振动轮激振频率相近发生共振。通过调整激振频率,降低了司机耳旁噪声。所得研究成果可为振动轮辐射噪声的预估与改进提供一种切实可行的参考依据。     

面向信息化战争的广义兰切斯特战争模型

21世纪是一个信息化时代,为了更好的面向信息化战争.通过对传统的兰切斯特微分方程进行改进,考虑了信息化的诸多因素,引入了战场情报感知系数和信息优势系数.构建了数学模型,并对其进行差分处理求解,比较分析了交战双方对战场感知系数变化和信息优势系数变化情况下,对战争的影响.以海湾战争为例,检验了模型的可行性.     

Sensory Perception of Non-Deuterated and Deuterated Organic Compounds

The chemical background of olfactory perception has been subject of intensive research, but no available model can fully explain the sense of smell. There are also inconsistent results on the role of the isotopology of molecules. In experiments with human subjects it was found that the isotope effect is weak with acetone and D₆-acetone. In contrast, clear differences were observed in the perception of octanoic acid and D₁₅-octanoic acid. Furthermore, a trained sniffer dog was initially able to distinguish between these isotopologues of octanoic acid. In chromatographic measurements, the respective deuterated molecule showed weaker interaction with a non-polar liquid phase. Quantum chemical calculations give evidence that deuterated octanoic acid binds more strongly to a model receptor than non-deuterated. In contrast, the binding of the non-deuterated molecule is stronger with acetone. The isotope effect is calculated in the framework of statistical mechanics. It results from a complicated interplay between various thermostatistical contributions to the non-covalent free binding energies and it turns out to be very molecule-specific. The vibrational terms including non-classical zero-point energies play about the same role as rotational/translational contributions and are larger than bond length effects for the differential isotope perception of odor for which general rules cannot be derived.     

基于双向循环神经网络的汉语语音识别*

当前基于深度神经网络模型中,虽然其隐含层可设置多层,对复杂问题适应能力强,但每层之间的节点连接是相互独立的,这种结构特性导致了在语音序列中无法利用上下文相关信息来提高识别效果,而传统的循环神经网络虽然做出了改进,但是只能对上文信息进行利用。针对以上问题,该文采用可以同时利用语音序列中上下文相关信息的双向循环神经网络模型与深度神经网络模型相结合,并应用于语音识别。构建具有5层隐含层的模型,其中第3层为双向循环神经网络结构,其他层采用深度神经网络结构。实验结果表明:加入了双向循环神经网络结构的模型与其他模型相比,较好地提高了识别正确率;噪声对双向循环神经网络汉语识别有重要影响,尤其是训练集和测试集附加噪声类型不同时,单一的含噪声语音的训练模型无法适应不同噪声类型的语音识别;调整神经网络模型中隐含层神经元数量后,识别正确率并不是一直随着隐含层中神经元数量的增加而增加,神经元数量数目增加到一定程度后正确率出现了降低的趋势。     

Computational gibberellin‐binding channel discovery unraveling the unexpected perception mechanism of hormone signal by gibberellin receptor

Gibberellins (GAs) are phytohormones essential for many developmental processes in plants. In this work, fundamental mechanism of hormone perception by receptor GID1 has been studied by performing computational simulations, revealing a new GA‐binding channel of GID1 and a novel hormone perception mechanism involving only one conformational state of GID1. The novel hormone perception mechanism demonstrated here is remarkably different from the previously proposed/speculated mechanism [Murase et al., Nature 2008 , 456, 459] involving two conformational states (“OPEN” and “CLOSED”) of GID1. According to the new perception mechanism, GA acts as a “conformational stabilizer,” rather than the previously speculated “allosteric inducer,” to induce the recognition of protein DELLA by GID1. The novel mechanistic insights obtained in this study provide a new starting point for further studies on the detailed molecular mechanisms of GID1 interacting with DELLA and various hormones and for mechanism‐based rational design of novel, potent growth regulators that target crops and ornamental plants. © 2013 Wiley Periodicals, Inc.     

基于感知的右视点质量可分级编码算法

为了减少立体视频传输如此庞大的数据量，根据人眼的立体掩蔽效应，提出了基于感知的右视点质量可分级编码算法．该算法将宏块分为3类：平坦块、纹理块和边缘块，通过不同的量化步长使得右视点图像块的质量可分级．实验证明根据左视点的图像质量，右视点在人眼感知范围内码率减少25％～37％．该算法能在人眼主观察觉范围内很好地提高码率，进一步消除人眼立体视觉冗余．