Deep Learning Tutorial By Hung-yi Lee
2017-02-10 17:33:38 0 举报AI智能生成
《深度学习教程》是由Hung-yi Lee编写的一本关于深度学习的经典教材。本书详细介绍了深度学习的基本概念，包括神经网络、反向传播算法、优化方法等，并通过大量的实例和代码，帮助读者深入理解和掌握深度学习的原理和应用。此外，本书还涵盖了深度学习的最新研究进展，如卷积神经网络、循环神经网络、深度强化学习等。无论你是深度学习的初学者，还是有一定基础的研究者，都能在这本书中获得宝贵的知识和启示。
作者其他创作
大纲/内容
Introduction of Deep Learning
Introduction of Deep Learning
General Machine Learning
NN
Output Layer
Goodness of function
Pick the best function
Gradient Descent
Why Deep not Fat?
A Example
"Hello World" for Deep Learning
Theano
learn 1
learn 2
Keras
Document
Examples
Example： Handwriting Digit Rec
Define a set of function
goodness of function
pick the best function
save and load models
use
Use GPU
THEANO_FLAGS=device=gpu0 python YourCode.py
import os
os.environ["THEANO_FLAGS"] = "device=gpu0"
os.environ["THEANO_FLAGS"] = "device=gpu0"
Tips for Training Deep Neural Network
Do not always blame Overfitting
Different approaches for different problems.
Choosing proper loss
Mini-batch
New activation function
Deeper usually does not imply better
In 2006, people used RBM pre-training.
In 2015, people use ReLU(Rectified Linear Unit)
Maxout
Adaptive Learning Rate
Popular & Simple Idea: Reduce the learning rate by some factor every few epochs.
Adagrad
Momentum
子主题
Problem:Overfitting
Why
Learning target is defined by the training data.
Panacea
Have/Create more training data
Methods
Early Stopping
Weight Decay
Dropout
Training with dropout, test without dropout
Dropout is a kind of ensemble
Ensemble
Dropout
Try
Network Structure
TASK
Variants of Neural Network
CNN: Convolutional NN
Why for Image?
1st layer (whole image) is too large
Some patterns are much smaller than the whole image
The same patterns appear in different regions
Subsampling is enough
What is CNN
tips
靠Convolution和Max Pooling降维
可以用多个filter进行分别做卷积
最后flatten
RNN: Recurrent NN
Next Wave
0 条评论
下一页
