Day 1
lecture 1: Introduction to ML and review of linear algebra, probability, statistics (kai)
lecture 2: linear model (tong)
lecture 3: overfitting and regularization (tong)
lecture 4: linear classification (kai)
Day 2
lecture 5: basis expansion and kernel methods (kai)
lecture 6: model selection and evaluation (kai)
lecture 7: model combination (tong)
lecture 8: boosting and bagging (tong)
Day 3
lecture 9: overview of learning theory (tong)
lecture 10: optimization in machine learning (tong)
lecture 11: online learning (tong)
lecture 12: sparsity models (tong)
Day 4
lecture 13: introduction to graphical models (kai)
lecture 14: structured learning (kai)
lecture 15: feature learning and deep learning (kai)
lecture 16: transfer learning and semi supervised learning (kai)
Day 5
lecture 17: matrix factorization and recommendations (kai)
lecture 18: learning on images (kai)
lecture 19: learning on the web (tong)
lecture 20: summary and road ahead (tong)
本文中讲的model是指的一个learning algorithm,甚至比learning algorithm所指的范围还要小,因为在一个learning algorithm里,不同的参数调节和不同的输入特征都会导致不同的model。模型选择的目标是使模型有更好的可解释性和更好的性能,而模型平均的目标只需要使模型有更好的性能即可,因为模型平均过程中用到了很多模型,而模型个数越多则其可解释性就越低。模型平均的英文名称有model ensemble,model blending, model combination, model averaging.
Model selection 和 model combination的不同使用体现在,如果某个模型以绝对的优势好于其他所有模型,那么这时候我们就采用model selection,因为不仅有好的性能,还可以获得好的可解释性。如果所有的模型在性能表现上都差不多,没有所谓的好坏,且模型本身又有很大的不同,这时候就可以采用model combination来大大提高其性能了。通常来说,model combination比model selection要稳定些。