Finite-size effects in separable recurrent neural networks