Supervised Machine Learning Regression and Classification #1

Machine Learning Specialization

• Supervised Machine Learning: Regression and Classification
• Advanced Learning Algorithms
• Unsupervised Learning, Recommenders, Reinforcement Learning

 

Supervised Machine Learning: Regression and Classification [Course]

Week 1 정리!!

Machine Learing

• Supervised Learning(Labels)
  • Regression (e.g. more experiences abroad ⇒ language skills)
  • classification (e.g. malevolent vs benign tumors)
• Unsupervised Learning(Without Labels/ algorithm has to find structure in the data)
  • clustering(e.g Google News)
    • group similar data points together
  • dimensionality reduction
    • compress data using fewer numbers
  • anomaly detection
    • find unusual data point

 

Week 1 정리!!

Machine Learing

• Supervised Learning(Labels)
  • Regression (e.g. more experiences abroad ⇒ language skills)
  • classification (e.g. malevolent vs benign tumors)
• Unsupervised Learning(Without Labels/ algorithm has to find structure in the data)
  • clustering(e.g Google News)
    • group similar data points together
  • dimensionality reduction
    • compress data using fewer numbers
  • anomaly detection
    • find unusual data points

 

Regression

• Goal: Finding Parameter ($w, b$) that minimizes Cost Function min J($w,b$)
• Gradient Descent: Algorithm that automatically finds best fit line that minimizes the cost function
  • Goal: Find min J($w,b$) as efficiently as possible
    1. start with some w, b (set w=0, b=0 per se)
    2. from that point take a look around 360, take a baby step in the direction has the steepest descent
    3. repeat step 2 until you cannot go down anymore (you found yourself a local minima, but is it the global minimum?)
    • if you start at different point, you might end up at different local minima!

(key point: w, b should be simultaneously updated)

 

derivative term을 이해하기 위한 2차원적 표현(parameter가 w 하나라고 가정)

• derivative term이 포인트에서 내려가게 만드는 핵심적인 부분
• learning rate은 얼마나 보폭을 크게 움직일 것인지 결정

 

• learning rate이 크면 , derivate term이 커져서 저렇게 계속 더 벌어질 수 있음!! (Fails to converge, may even diverge)

• local minimum에 다가갈수록 derivative term이 작아지면서 update 보폭도 점점 줄어듦!
• 그래서 fixed learning rate을 사용할 수 있는 거임!

 

For Linear Regression

 

• 각 gradient descent step에서 총 cost를 계산하기 위해 모든 training data point를 사용하는 것을 “Batch” gradient descent라고 함!