2022秋week1，9月12日
作者：互联网
上海—小雨
2022秋week1
9月12日
计算机视觉：
why cv matters？
for safety health security comfort fun access and so on.
course contects:
- salieny detection
- segmentation
- object detection
- object recognition
- image recognition
- video processing
categories for ML
深度学习：
lecture notes 01:
lecture logistic
intro to deep learning
machine learning review
Artificial neurons
General learning process
deep learning is using the deep neural network as the map function.
The Inspiration of using neural network to solve problem came from visual cortex. PAPER LINK
Deep network is more compactly and learn more representation of input data.
Math review
- Gradient
- Local and global minima
Necessary condition: The derivative is zero
Sufficient condition: Hessian is positive definte about Hessian matrix ,we can go to LINK and Taylor formula.
-
Probability chain rule:
-
Bayes rules:
-
common distribution:
高斯分布
Standard learning scenarios:
include unsupervised learning; supervised learning; semi-supervised learning; reinforcement learning
- supervised learning
Learning problem
Learning as iterative optimization
Supervised learning pipeline
train dataset训练parameter validation data 训练 hyper parameter
Generalization
Model selection for better generalization
Questions
生物神经元和计算神经元的比较
Capacity of single neuron
What a single neuron does?
激活函数是怎么起作用的？
经验损失：
因为是通过已有数据得到的损失和风险。
理论上对于已有数据，已经在统计学上进行了反复的计算理解，这些计算出的风险和损失是通过先前对数据理解的经验获得的，因此叫经验损失。
随机梯度下降：
机器学习-损失函数（0-1损失函数、绝对损失函数、平方损失函数、对数损失函数）
点到直线的距离公式
协方差矩阵
Single-Layer Neural Networks and Gradient Descent
argmax和argmin
arg min 就是使后面这个式子达到最小值时的变量的取值
arg max 就是使后面这个式子达到最大值时的变量的取值
负对数似然(negative log-likelihood)
机器学习-正则化
铰链损失函数（Hinge Loss）的理解 LINK1 LINK2
林轩田《机器学习基石》：https://github.com/RedstoneWill/HsuanTienLin_MachineLearning/tree/master/Machine%20Learning%20Foundations/pdf%20files
感知器算法（Perceptron Algorithm） 证明
不失一般性的（WLOG）
lecture notes 02:
标签：12,函数,损失,supervised,2022,learning,week1,lecture,network 来源： https://www.cnblogs.com/duzetao/p/16686577.html