Gradient descent

Gradient descent (in 3D matrix) in a nutshell:

From a random point in 3D, finding a bottom or bottoms, by gradually moving downward. 

How:

repeat {

     θj := θj − α (∂J(θ)/∂θj)

}    (simultaneously update for j = 0 and j = 1)

What it means:

If current standpoint is on / this shape of slope, go left and if \, go right.  

This is just common sense. Assume you are climbing down the mountain. It the slope is / shape, you go left, and go right otherwise. As simple as that.