前言

本筆記來自台大林軒田老師的CSIE5043 Machine Learning 2021FALL課程,於此做一個學習記錄。


What is Machine Learning

skill: improve some performance measure(e.g. prediction accuracy)

技巧是指提高效能觀測指標的能力,例如提高預測的精確度。

相對於人類從觀察進行技巧學習的模式,機器從data中進行技巧學習。


Why use ML

例如在樹木辨識這一議題中,倘若使用傳統的編程解決,那麼將面臨樹木這一概念難以定義的問題。而建構一個ML樹木辨識系統卻相對簡單的多。這意味著面對複雜問題時,ML是一種可選的替代方案。

Some Use Scenarios(ML應用場景):

when human cannot program the system manually.

e.g. navigating on Mars

when human cannot 'define the solution' easily.

e.g. speech/visual recognition

when needing rapid decisions that huamns cannot do.

e.g. high-frequency trading

when needing to be user-oriented in a massive scale

e.g. consumer-targeted marketing

Key Essence of ML(使用ML的要素)

  1. exists some 'underlying pattern' to be learned. So 'performance measure' can be improved
  2. but no programmable(easy) definition.
  3. somehow there is data about pattern. So ML has some 'inputs' to learn from

Application of ML

疾病診斷:

4G通訊:

缺陷PCB檢測:


Components of ML

Formalize the Learning Problem:

ML: use data to compute hypothesis that approximates target

ML:用資料集從潛在函數集合中選擇出最接近理想target function的一個 >


ML and other fields

  • ML and Data Mining

    ML: use data to compute hypothesis that approximates target

    DM: use (huge) data to find property that is interesting.

    在現實情況中ML和DM往往難以區分。

  • ML and Statistics

    ML: use data to compute hypothesis that approximates target

    Statistics: use data to make inference about an unknown process.

    現實中統計學為ML提供了許多可用的工具


Perceptron(感知器) Hypothesis Set

Perceptron is a simple hypothesis set.

對消費者的各種feature(年齡薪水等)設計不同的權重,當其feature加權和大於某門檻時則允許發卡,反之拒絕。

將hypothesis set中的函數h(x)繼續用數學推導簡化(將threshold歸納進權重矩陣):

不同的權重矩陣代表了不同的hypothesis

Perceptrons in :

隨著vector 的不同平面上的直線不同,對平面上的point(代表不同消費者)進行binary classification.

如何選擇接近target function 呢?這邊採用的策略是如下不斷修正,Perceptron Learning Algorithm(PLA):

迭代過程:

省略

PLA可行性證明:

因為所以得上式,接著得到下式一:

$$

$$

is inner product before updating, is inner product after updating. 在每次進行更新後,inner product都會增加。同時討論下式:

changed only when mistake

代入故有以下式二:

$$

$$

由式一式二累加相消且取為zero vector後得到:

即在線性可分的情況下,PLA進行mistake corrections的上限次數被約束在,證明PLA運作的可行性