📄 中文摘要
该研究提出了一种可并行化的神经图灵机（P-NTM）简化版本，通过重新设计原始架构的核心操作，实现高效的基于扫描的并行执行。对该架构在一系列算法问题的合成基准上进行了评估，这些问题涉及状态跟踪、记忆和基本算术，并通过自回归解码进行求解。与标准神经图灵机的稳定实现以及传统的递归和基于注意力的架构进行了比较。结果表明，尽管进行了简化，所提出的模型在长度泛化性能上与原始模型相当，能够学习解决所有问题，包括未见的序列长度。
可并行化的神经图灵机
出处: Parallelizable Neural Turing Machines
发布: 2026年2月24日
📄 English Summary
Parallelizable Neural Turing Machines
This study introduces a parallelizable simplification of the Neural Turing Machine (P-NTM), which redesigns the core operations of the original architecture to enable efficient scan-based parallel execution. The proposed architecture is evaluated on a synthetic benchmark of algorithmic problems involving state tracking, memorization, and basic arithmetic, solved via autoregressive decoding. Comparisons are made against a revisited stable implementation of the standard NTM, as well as conventional recurrent and attention-based architectures. Results indicate that, despite its simplifications, the proposed model achieves length generalization performance comparable to the original, successfully learning to solve all problems, including those with unseen sequence lengths.