Oja rule as PCA power iteration: w(t+1) = w(t) + eta*(y*x - y^2*w); w_inf = first eigenvector

Oja's rule as power iteration on the input covariance. Setup: a single linear neuron y = w*x receiving inputs x with zero mean and covariance C = E[x*x^T]. Oja 1982 modified Hebbian learning: delta-w = eta * (y*x - y^2*w), the y^2*w term…