Forward accessibility for recurrent neural networks