Actually, "learns" here means to fit reverse linear transformation a.k.a. inverse matrix. He defines reverse FT matrix and using gradient descent numerically co

Actually, "learns" here means to fit reverse linear transformation a.k.a. inverse matrix. He defines reverse FT matrix and using gradient descent numerically converges to inverse matrix. Nothing more than inefficient way to solve system of linear equations.