Take vague input data.
lay good fundation.
Talk is cheap, show me the code.
Background:
- Make decsion/prediction based on Data.
- data visualization and representation.
这次机器学习基础课程的重点仍然是有监督与无监督学习
Build models work for unseen data对尚未观测到的数据进行预测。
Terminology：
Types:
Supervised Learning: learn from training Data example->ml model :predict an unseen data.
data+ label
Self-supervided/transfer-learning
Key:supervised learning.
Traing example:training data and value.
目标，找到什么函数可以take x to y.
Training Data
Are feature informative?
Invariances:理解、建模和处理数据不变性是机器学习中非常重要的主题。
Error on small groups,label imbalance.
训练数据的一部分标签是你所不感兴趣的。
Outliers：regression
有的时候outlier也未必是坏事，有的时候离群值可以让模型变得更好。
signal or noise?
Representation of data
numerical
categorical
binary
structured
very deserve exercise
How to represent raw image.
Training Data and test data have same underlying distritutes:这样你的模型就会效果很好(small test error)。
也就是说，你想在一堆图片中找到橘子，你的训练和测试都有橘子。
Zero error?
data hungry model:complex model ,memory the train data
Overfitting and 鲁棒性
讨论分类的鲁棒性
优化: minimize loss 凸模型
Model Selection
How do I know my model makes sense?
which feature is influential?
- Simple vs complex :generalization
- Simple vs complex:interpretability
