- Reinforcement Learning with TensorFlow
- Sayon Dutta
- 82字
- 2025-02-23 17:35:22
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 ?