Training-based technology: as using data and training data, the machine itself can make own algorithm by finding the rules or patterns.
without being explicitly programmed
purpose of machine learning
finding the good (or optimized hypothesis) that the cost (or error) is low
but, before entering the data, the user has to first design the specific hypothesis when the user does not know about the formula.
to avoid the overfitting, regulation for the machine learning (by lessening the size of the weight) is needed. argmin = $ \begin{matrix} min & cost \\ weight(= \theta) & \end{matrix} $
How to train the machine?
calculate the value of cost (with the prediction using the first hypothesis set by the user) and then, differentiate the value of cost to find the slope.
the initial weight is for lowering the value of cost
W(t+1) <= Wt - a * Gradient (Repeating this process)
The dataset for machine learning
a part of dataset is for training, the rest of the dataset is for validation.
when the result of the training data is good but the result of validation is not, the training is useless.