[机器学习入门] 李宏毅机器学习笔记-10 （Tips for Deep Learning；深度学习小贴士）

PDFVIDEO

---

Recipe of Deep Learning

---

在 training data上的performance不好

---

Deeper usually does not imply better

---

Vanishing Gradient Problem

---

ReLU（Rectified Linear Unit）

ReLU - variant

**那么除了ReLU有没有别的activation function了呢？ 所以我们用 Maxout来根据training data自动生成activation function** ReLU is a special cases of Maxout

---

Maxout

---

ReLU is a special cases of Maxout

---

More than ReLU

---

Maxout - Training

---

Adaptive Learning Rate

---

RMSProp

---

Hard to find optimal network parameters

---

Momentum（gradient descent 融入惯性作用）

所以，加了momentum后：

---

Adam

---

在testing data上performance不好

---

Early Stopping

---

Regularization

---

Regularization - Weight Decay

---

Dropout

---

Dropout- Intuitive Reason

---

Dropout is a kind of ensemble