2018-07-31 17:22:40 lz_peter 阅读数 854

Lecture I : Introduction of Deep Learning

• Introduction of Deep Learning
• Step1 : define a set of function
• Step2 : goodness of function
• Step3 : pick the best function
• Why Deep?
• "Hello World" for Deep Learning

Lecture II : Tips for Training Deep Neural Network

• Recipe of Deep Learning
• Choosing proper loss
• Mini-batch
• New activation function
• Momentum
• Early Stopping
• Weight Decay
• Regularization
• Dropout
• Network Structure

Lecture III : Variants of Neural Network

• Convolutional Neural Network(CNN)
• Recurrent Neural Network(RNN)

Lecture IV : Next Wave

• Supervised Learning
• Ultra Deep Network
• Attention Model
• Reinforcement Learning
• Unsuperivised Learning
• Image:Realizing what the World Looks Like
• Text:Understanding the Meaning of Words
• Audio:Learning human language without supervision

2019-11-27 16:14:40 wozaipermanent 阅读数 26

# 1 Introduction of Deep Learning

## 1.1 Three Steps for Deep Learning

• Step1: define a set of function (Neural Network)
• Step2: goodness of function
• Step3: pick the best function

## 1.2 Step1: Neural Network

### 1.2.2 Output Layer(Option)

• Softmax(归一化指数函数)：它能将一个含任意实数的k维向量Z“压缩”到另一个k维向量$\sigma(Z)$中，使得每一个元素的范围都在(0, 1)之间，并且所有元素的和为1。

### 1.2.3 Example Application

• Handwriting Digit Recognition

• Total Loss:

## 1.4 Step3: Pick the Best Function

• RBM(Restricted Boltzmann Machine): 受限玻尔兹曼机，这部分可以参考链接：https://zhuanlan.zhihu.com/p/22794772

• Then Compute $\partial L / \partial w$ , if Negative then Increase w; elif Positive then decrease w

• $\eta$ is called “learning rate”

• Randomly pick a starting point

• Backpropagation(反向传播算法)：an efficient way to compute $\partial L / \partial w$ , link below:

## 1.5 Deep is Better

### 1.5.2 Thin + Tall is Better

• Neural network consists of neurons

• A hidden layer network can represent any continuous function

• Using multiple layers of neurons to represent some functions are much simper

• Less parameters, less data

## 1.6 Toolkit

### 1.6.2 Example of Handwriting Digit Recognition

#### Testing

score = model.evaluate(x_test, y_test)
print('Total loss on Testing Set: ', score[0])
print('Accuracy of Testing Set: ', score[1])
result = model.predict(x_test)


### 1.6.3 GPU to Speeding Training

• Way1

THEANO_FLAGGS=device=gpu0 python YourCode.py

• Way2

import os
os.environ["THEANO_FLAGS"] = "device=gpu0"


# 2 Tips for Training Deep Neural Network

## 2.1 Good Results on Training Data

### 2.1.3 New Activation Function

#### ReLU

model.add(Activation('sigmoid'))


#### Learning Rates

• If learning rate is too large, total loss may not decrease after each update
• If learning rate is too small, training would be too slow

Notes:

• Learning rate is smaller and smaller for all parameters
• Smaller derivatives, larger learning rate, and vice versa

## 2.2 Good Results on Testing Data

### 2.2.1 Early Stopping

#### Why Overfitting

• Learning target is defined by the training data.
• The parameters achieving the learning target do not necessary have good results on the testing data.

### 2.2.2 Weight Decay

Weight decay is one kind of regularization.

• Our brain prunes out the useless link between neurons.
• Doing the same thing to machine’s brain imporves the performance.

### 2.2.3 Dropout

#### Training

• Each time before updating the parameters

• Each neuron has p% to dropout
• The structure of the network is changed.
• Using the new network for training
• For each mini-batch, we resample the dropout neurons

### 2.2.4 Network Structure

e.g. CNN is another good example.

# 3 Variants of Neural Network

## 3.1 Convolutional Neural Network (CNN)

### 3.1.1 Why CNN for Image

• When processing image, the first layer of fully connected network would be very large.
• Some patterns are much smaller than the whole image. A neuron does not have to see the whole image to discover the pattern.
• The same patterns appear in different regions.
• Subsampling the pixels will not change the object, so we can subsample the pixels to make image smaller.

### 3.1.2 Three Steps

#### Step1: Convolutional Neural Network

##### Max Pooling

• Smaller than the original image.
• The number of the channel is the number of filters.

# 4 Next Wave

## 4.2 Reinforcement Learning

### 4.2.3 Difficulties of Reinforcement Learning

• It may be better to sacrifice immediate reward to gain more long-term reward.
• Agent’s actions affect the subsequent data it receives.

## 4.3 Unsupervised Learning

### 4.3.2 Text: Understanding the Meaning of Words

• Machine learn the meaning of words from reading a lot of documents without supervision
• A word can be understood by its context

### 4.3.3 Audio: Learning Human Language Without Supervision

• Audio segment corresponding to an unknown word (Fixed-length vector)
• The audio segments correspondsing to words with similar pronunciations are close to each other.

2018-11-18 16:00:33 weixin_41913844 阅读数 211

Deep Learning
• 上學期的「機器學習」錄影
• Deep generative model (Part 1):
• Deep generative model (Part 2):

2018-11-05 09:28:36 fengdu78 阅读数 9

1.讲义大纲

1.1   深度学习概论（p5）

1.1.1 深度学习的三个步骤（p10）

1）定义一系列函数（p11）

1.1.2 函数的优点（p26）

1） 训练数据（p27）

2）学习目标（p28）

3）损失函数（p29）

1.1.3 选择最佳函数（p32）

1）梯度下降（p33）

2）反向传播推导（p44）

1.2   为什么使用深度（p47）

1.2.1 更多参数，更优性能（p47）

1.2.2  任何函数可以通过一个单一的隐藏层实现（p44）

1.2.3  深度学习：模块化？需要更少的数据（p52）

2.1 合适的损失函数（p69）

2.1.1 平方误差和交叉熵（p74）

2.2 Mini batch （p74）

2.2.1 更好的性能（p83）

2.3 激活函数（p86）

2.3.1 RELU（p92）

2.3.2 Maxout（p98）

2.4 调整学习率（p98）

2.5 Momentum（p108）

2.6 解决过拟合（p115）

2.6.1 更多的训练数据（p116）

2.7 早停（p119）

2.8 权重衰减（p121）

2.9 Dropout（p126）

2.10 网络架构（p138）

3.1 CNN （p149）

3.1.1 卷积（p158）

3.1.2 池化（p165）

3.1.3 平铺（p170）

3.2 RNN（p192）

3.3 LSTM（p196）

3.4 GRU（p211）

4.1 监督学习（p226）

4.1.1 超级深的网络（p226）

4.1.2 注意力模型（p235）

4.2 增强学习（p252）

4.3 无监督学习（p264）

2.讲义截图：

3.总结：

QQ群：654173748

2017-08-04 21:44:31 soulmeetliang 阅读数 2583

[机器学习入门] 李宏毅机器学习笔记-37(Deep Reinforcement Learning;深度增强学习入门)

PDF VIDEO

# Deep Reinforcement Learning

## Scenario of Reinforcement Learning

### Learning to paly Go

AlphaGo 采取的策略是先用监督学习learn的不错后，再用增强学习狂下棋。

### Example: Playing Video Game

Play yourself: http://www.2600online.com/spaceinvaders.htm l • How

## Outline

Alpha Go 用的方法是：policy-based + value-based + model-based

## Policy-based Approach

Learning an Actor

function是Pi，input是environment，output是Action。

### Step 1： Neural Network of function

NN的好处就是比较generalized，即使没见过的场景，也可能给出合理的结果。

### Step 2:goodness of function

Review: Supervised learning

### Step 3: pick the best function

The probability of the actions not sampled will decrease.

## Value-based Approach

Learning a Critic

end！