Reinforcement Learning with TensorFlow
上QQ阅读APP看书，第一时间看更新
上一章目录下一章

Overcoming the limitations of deep learning

These two possible problems can be overcome by:

  • Minimizing the use of the sigmoid and tanh activation functions
  • Using a momentum-based stochastic gradient descent
  • Proper initialization of weights and biases, such as xavier initialization
  • Regularization (add regularization loss along with data loss and minimize that)
For more detail, along with mathematical representations of the vanishing and exploding gradient, you can read this article: Intelligent Signals : Unstable Deep Learning. Why and How to solve them ?
上一章目录下一章