Reservoir stack machines

Memory-augmented neural networks equip a recurrent network with an explicit memory to support tasks that require information storage without interference over long times. A key motivation for such research is perform classic computation tasks, as parsing. However, memory-augmented are notoriously hard train, requiring many backpropagation epochs and lot of data. In this paper, we introduce the reservoir stack machine, model which can provably recognize all deterministic context-free languages circumvents training problem by only output layer net employing auxiliary during about desired interaction stack. our experiments, validate machine against deep shallow from literature on three benchmark Neural Turing machines six languages. Our results show achieves zero error, even test sequences longer than data, few seconds time 100 sequences.