首页> 外文会议>International Conference on Robot Intelligence Technology and Applications >Learning Time Constant of Continuous-Time Neurons with Gradient Descent
【24h】

Learning Time Constant of Continuous-Time Neurons with Gradient Descent

机译:与梯度下降的连续时间神经元的学习时间常数

获取原文

摘要

In this paper, we propose a learning method to update the time constant in each continuous-time neuron with gradient descent to generate desired output patterns. Selecting appropriate time constant for each neuron in a continuous-time recurrent neural network is difficult. Hence, the development of adaptive method of the time constant is desired. However, direct update of time constants with gradient descent is significantly unstable. Therefore, to avoid the instability, we propose a learning method applying gradient descent to the logarithm of the time constant. We carried out an oscillator reproducing task in which a learning network is trained to generate the same oscillatory outputs from the teacher network. The training result shows that our proposed method can successfully update the time constants and suggests that leaning of time constants expands the freedom in learning and improve the learning performance.
机译:在本文中,我们提出了一种学习方法,以利用梯度下降来更新每个连续时间神经元的时间常数,以产生所需的输出模式。在连续时间复发神经网络中为每个神经元选择适当的时间常数是困难的。因此,期望开发时间常数的自适应方法。但是,直接更新具有梯度下降的时间常数是显着不稳定的。因此,为了避免不稳定,我们提出了一种学习方法将梯度下降应用于时间常数的对数。我们执行了一个振荡器再现任务,其中培训学习网络以产生来自教师网络的相同振荡输出。培训结果表明,我们的提出方法可以成功更新时间常数并表明倾斜时间常数扩大了学习自由和提高学习绩效。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号