官术网_书友最值得收藏!

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 ?
主站蜘蛛池模板: 涿鹿县| 独山县| 富源县| 尚志市| 泰宁县| 海淀区| 莲花县| 雷山县| 洛隆县| 恩平市| 达拉特旗| 称多县| 西乡县| 屏边| 宣威市| 新余市| 灵台县| 黑龙江省| 五指山市| 北票市| 水城县| 肥东县| 图木舒克市| 尼玛县| 聊城市| 类乌齐县| 禹城市| 股票| 绥阳县| 新源县| 博客| 红桥区| 台北市| 永靖县| 哈巴河县| 九江市| 凌海市| 响水县| 进贤县| 克山县| 富锦市|