I’m trying to implement backpropagation from scratch. I’m trying to update the weights of a network with multiple inputs and multiple outputs but no hidden layers. The logic I have implemented seems fine to me. But the weights are not converging but rather increasing. I’m not able to debug as to why it is happening.

def update():
    global weights_n
    for i in range(10000):
        y_mult = p1(x_random)
        print("Varying Output",y_mult)
        print(weights_n)
        error = calculate_errors(y_mult, y)
        z=y_mult-y
        print("Error",error)
        grad_a1 = np.dot(x_random.T,z)
        print(learning_rate * grad_a1)
        weights_n = weights_n - learning_rate * grad_a1
        print(weights_n)
        if error.all()<=0.5:
            break
    return weights_n

When I try to use it with numpy sum and matmul, it works but with dot product something is off.