02. Descending into Machine Learning

Descending into ML

---

Topic: Descending into ML

Course: GMLC

Date: 10 February 2019 

Professor: Not specified

---

Resources

---

• https://developers.google.com/machine-learning/crash-course/descending-into-ml/video-lecture

• https://developers.google.com/machine-learning/crash-course/descending-into-ml/linear-regression

• https://developers.google.com/machine-learning/crash-course/descending-into-ml/training-and-loss

• https://developers.google.com/machine-learning/crash-course/descending-into-ml/check-your-understanding

Key Points

---

• Linear relationship

  • 

  • y' - the predicted label or a desired output

  • b - the bias or the y intercept (offset from an origin)

  • w1 - weight of feature 1 (x1) (same as slope m from traditional line equation) 

  • x1 - feature or a known input

  • The following example uses only one feature, but model may use multiple features and weights

• Linear regression - type of regression model that produces a continuous value from a linear combination of features

• Inference - making the prediction itself using the unlabelled examples (after training the model with labeled examples)

• Weight - a coefficient for a specific feature, or an edge in a deep network

• Training - determining good values for all the weights and bias from labeled examples by examining models to find one that minimises the loss (process called empirical risk minimisation )

• Loss - the number indicating how bad the prediction was on a simple example (perfect - 0, only greater than 0)

  • Goal of a good prediction model is to have an average low loss

  • https://www.geogebra.org/graphing/jjqtmmad

• Squared loss

  •  the square difference between the label and the prediction

  • (observation - prediction(x))2 = (y-y’)2

• Mean square error (MSE)

  • 

  • Average squared loss for each example over the whole data set

  • x - set of features (numeric value)

  • y - label (for example a temperature)

  • prediction(x) - function of weights and bias in combination with the features themselves

  • D - Data set combing many (x, y) labeled pairs / examples

  • N - Number of examples in data set

Check your understanding

---

• Know the variable names for machine learning functions

• Know how do weight and feature correlate with each other

• Know how to explain empirical risk minimisation and where is it used

• Know how do calculate loss

• Understand the MSE calculations and the quality of prediction model by looking at MSE results

Summary of Notes

---

• Model’s quality is determined by its MSE result

• Empirical risk optimisation is a training process in which models are examined to find the one which has the lowest average loss