Anatomy of Associative Recall in Fixed-State Recurrences: A Matched-State Decomposition, an Interference Wall, and a Curriculum That Breaks It
本文分解了固定状态循环(fixed-state recurrences)与注意力机制在联想记忆(associative recall)任务上的性能差距,揭示了缺失的卷积(missing convolutions)和训练干扰(training interference)而非固有的架构限制才是导致该差距的主要原因,并证明了针对性的距离课程学习(distance curriculum)可以可靠地克服这些障碍,从而实现完美的记忆。