Natural Gradient Step

Problem

Perform one natural gradient descent step with a diagonal Fisher information matrix:

params_new[i] = params[i] - eta * gradient[i] / fisher_diag[i]

Return the updated parameter list.

Hint

Dividing by the Fisher information rescales steps by parameter uncertainty.

case 1

in: [[0,0],[1,2],[1,1],0.1]

expected: [-0.1,-0.2]

case 2

in: [[1,-1],[0.5,-0.5],[2,2],0.5]

expected: [0.875,-0.875]

case 3

in: [[0],[1],[4],1]

expected: [-0.25]

case 4

in: [[2,3],[0,0],[1,1],0.1]

expected: [2,3]

Starter codePython

def natural_gradient_step(params, gradient, fisher_diag, eta):
    # Your code here
    pass

4 test cases · run your code against them in the editor