机器学习的目标
- Regression
- Classification
- Stuctured Learning
Machine Learning <=> Look for function
- Function with unknown parameter (Model)
- Define loss from train data
- input: w,b
- output: how good a set of value is
- Optimization w*, b* = argmin L
Piecewise linear
From the perspective of deep learning, all curve can be approximated by constant + sum of a set of hard sigmoid
c sinmoid(b + wx)
- w: change slopes
- b: shigt
- c : height
When x is single variable
- y = b + sigma_i (c_i * sigmoid(b_i + w_i * x_i)
When x is multivariate
- y = b + sigma_i (c_i * sigmoid(b_i + sigma_j(w_ij * x_j))
that is: y = b + c^T sigmoid(b + w^T *X)
深度学习的目标
Based on the above piecewise linear, the goal of deep learning can be described below:
- Define function with unknown parameters: y = b + c^T sigmoid( b + w^T * X)
- Define loss function: L(theta)
- Optimization: theta* = argmin L(theta)
Sigmoid can also be replaced by ReLu (rectified linear unit)
Tow ReLu can synthesize the hard sigmoid