Lecture 22: Gradient Descent: Downhill to a Minimum
Description
Gradient descent is the most common optimization algorithm in deep learning and machine learning. It only takes into account the first derivative when performing updates on parameters—the stepwise process that moves downhill to reach a local minimum.
Summary
Gradient descent: Downhill from \(x\) to new \(X = x - s (\partial F / \partial x)\)
Excellent example: \(F(x,y) = \frac{1}{2} (x^2 + by^2)\)
If \(b\) is small we take a zig-zag path toward (0, 0).
Each step multiplies by \((b - 1)/(b + 1)\)
Remarkable function: logarithm of determinant of \(X\)
Related section in textbook: VI.4
Instructor: Prof. Gilbert Strang
Problems for Lecture 22
From textbook Section VI.4
1. For a 1 by 1 matrix in Example 3, the determinant is just \(\det X=x_{11}\). Find the first and second derivatives of \(F(X)=-\log(\det X)=-\log x_{11}\) for \(x_{11}>0\). Sketch the graph of \(F=-\log x\) to see that this function \(F\) is convex.
6. What is the gradient descent equation \(\boldsymbol{x}_{k+1}=\boldsymbol{x}_k-s_k\nabla f(\boldsymbol{x}_k)\) for the least squares problem of minimizing \(f(\boldsymbol{x})=\frac{1}{2}||A\boldsymbol{x}-\boldsymbol{b}||^2\)?
