林軒田-機器學習基石 課堂筆記（一）A takes D and H to get g

1.From Learning to Machine Learning

Learning: Observations->learning->skill

Machine Learning: data->ML->skill

ps: skill — improve some performance measure (eg: prediction accuracy)

2.Key Essence of ML:(help decide whether to use ML)

  ①exists some ‘underlying pattern’ to be learned

      —— so ‘performance measure’ can be improved

  ②but no programmable (easy) definition

      —— so ML is needed

  ③somehow there is data about the pattern

      —— so ML has some ‘input’ to learn from

3.The Learning Model（※）

A more appropriate definition of ML: use data to compute hypothesis g target f .

NOTE：課中以銀行是否會同意用戶申請辦理信用卡爲例

1>收集的（單個）用戶信息包含的內容，可理解爲輸入內容：

2>基本的符號含義以及其對應本例的含義：

3>假設函數集

從圖中我們可以知道機器學習從由大量數據分析中得到假設函數集，從而得到一個能近似表示輸入輸出集映射關係的目標函數g，通過g來進行預測，

更通俗的來講，機器學習是人工智能的一個分支，我們使用計算機設計一個系統，使他能根據提供的訓練數據按照一定的方法來學習，隨着訓練數據次數的增加，該系統可以在性能上不斷學習和改進，通過參數優化和學習模型，能夠用於預測相關問題的輸出。

4.ML vs DM/AI/Statistic

  ①Machine Learning: use data to compute hypothesis g that approximates target f

  ②Data Mining: use (huge) data to find property that is interesting

     PS: a. if ‘interesting property’ same as ‘hypothesis that approximate target’,

               ML=DM

            b. if ‘interesting property’ related to ‘hypothesis that approximate target’,

               DM can help ML (often, but not always)

  ③Artificial Intelligence: compute something that shows intelligent behavior

     PS: g ≈ f is something that shows intelligent behavior

           —— ML can realize AI, among other route

           e.g.: chess playing

                     traditional AI: game tree

                     ML for AI: ‘learning from board data’

  ④use data to make inference about an unknown process

     PS: g is a inference outcome;

           f is something unknown.

           —— Statistic can be used to achieve ML, traditional Statistic also focus on

        provable results with math assumptions, and care less about computation.