针对深度学习应用技术进行了研究性综述。详细阐述了RBM(受限玻尔兹曼机)逐层预训练后再用BP(反向传播)微调的深度学习贪婪层训练方法,对比分析了BP算法中三种梯度下降的方式,建议在线学习系统采用随机梯度下降,静态离线学习系统采用随...针对深度学习应用技术进行了研究性综述。详细阐述了RBM(受限玻尔兹曼机)逐层预训练后再用BP(反向传播)微调的深度学习贪婪层训练方法,对比分析了BP算法中三种梯度下降的方式,建议在线学习系统采用随机梯度下降,静态离线学习系统采用随机小批量梯度下降;归纳总结了深度学习深层结构特征,并推荐了目前最受欢迎的五层深度网络结构设计方法。分析了前馈神经网络非线性激活函数的必要性及常用的激活函数优点,并推荐Re LU(rectified linear units)激活函数。最后简要概括了深度卷积神经网络、深度递归神经网络、长短期记忆网络等新型深度网络的特点及应用场景,归纳总结了当前深度学习可能的发展方向。展开更多
为了减小训练集中各类别资源分布不均衡对分类性能造成的影响,该文对原始训练集使用类别均衡法,即对原始训练集以类为单位进行重新组合,使得重组后的训练集类别分布尽可能均衡,从而可以在均衡的类别上进行训练和分类,以降低在训练过程...为了减小训练集中各类别资源分布不均衡对分类性能造成的影响,该文对原始训练集使用类别均衡法,即对原始训练集以类为单位进行重新组合,使得重组后的训练集类别分布尽可能均衡,从而可以在均衡的类别上进行训练和分类,以降低在训练过程中对小类别的不公平待遇。在复旦大学语料库上使用类别均衡法,分别用N a ve B ayes和R occh io方法分类,前者的宏平均F1从48.62%提高到了80.99%,后者的宏平均F1从64.58%提高到80.26%,微平均F1从73.99%提高到80.47%。实验结果显示,类别均衡法显著提高了分类性能。展开更多
A three-dimensional pharmacophore model was developed from a series of inhibitors of Aurora A kinase to discover new potent anti-cancer agents using the HypoGen module in the Catalyst software. The pharmacophore model...A three-dimensional pharmacophore model was developed from a series of inhibitors of Aurora A kinase to discover new potent anti-cancer agents using the HypoGen module in the Catalyst software. The pharmacophore model was developed based on the structure of 20 currently available inhibitors, which were carefully selected from the literature. The best hypothesis （Hypo 1） was defined by four features： one hydrogen-bond donor and three hy- drophobic points, with the best correlation coefficient of 0.909, the lowest rms deviation of 1.563, and the highest cost difference of 99.075. The Hypo 1 was then validated by a test set consisting of 24 compounds and by a cross-validation of 95% confidence level through randomizing the data using the CatScramble program, which suggested that a predictive pharmacophore model had been successfully obtained.展开更多
文摘针对深度学习应用技术进行了研究性综述。详细阐述了RBM(受限玻尔兹曼机)逐层预训练后再用BP(反向传播)微调的深度学习贪婪层训练方法,对比分析了BP算法中三种梯度下降的方式,建议在线学习系统采用随机梯度下降,静态离线学习系统采用随机小批量梯度下降;归纳总结了深度学习深层结构特征,并推荐了目前最受欢迎的五层深度网络结构设计方法。分析了前馈神经网络非线性激活函数的必要性及常用的激活函数优点,并推荐Re LU(rectified linear units)激活函数。最后简要概括了深度卷积神经网络、深度递归神经网络、长短期记忆网络等新型深度网络的特点及应用场景,归纳总结了当前深度学习可能的发展方向。
文摘为了减小训练集中各类别资源分布不均衡对分类性能造成的影响,该文对原始训练集使用类别均衡法,即对原始训练集以类为单位进行重新组合,使得重组后的训练集类别分布尽可能均衡,从而可以在均衡的类别上进行训练和分类,以降低在训练过程中对小类别的不公平待遇。在复旦大学语料库上使用类别均衡法,分别用N a ve B ayes和R occh io方法分类,前者的宏平均F1从48.62%提高到了80.99%,后者的宏平均F1从64.58%提高到80.26%,微平均F1从73.99%提高到80.47%。实验结果显示,类别均衡法显著提高了分类性能。
基金Supported by the National Natural Science Foundation of China under Grant Nos.60702033 60772076 (国家自然科学基金)+3 种基金the National High-Tech Research and Development Plan of China under Grant No.2007AA01Z171 (国家高技术研究发展计划(863)the Science Fund for Distinguished Young Scholars of Heilongjiang Province of China under Grant No.JC200611 (黑龙江省杰出青年科学基金)the Natural Science Foundation of Heilongjiang Province of China under Grant No.ZJG0705 (黑龙江省自然科学重点基金)the Foundation of Harbin Institute of Technology of China under Grant No.HIT.2003.53 (哈尔滨工业大学校基金)
文摘A three-dimensional pharmacophore model was developed from a series of inhibitors of Aurora A kinase to discover new potent anti-cancer agents using the HypoGen module in the Catalyst software. The pharmacophore model was developed based on the structure of 20 currently available inhibitors, which were carefully selected from the literature. The best hypothesis （Hypo 1） was defined by four features： one hydrogen-bond donor and three hy- drophobic points, with the best correlation coefficient of 0.909, the lowest rms deviation of 1.563, and the highest cost difference of 99.075. The Hypo 1 was then validated by a test set consisting of 24 compounds and by a cross-validation of 95% confidence level through randomizing the data using the CatScramble program, which suggested that a predictive pharmacophore model had been successfully obtained.
