把 CTC loss 的激活内存从词表线性增长压下来,同时保持 loss 与一阶梯度等价,直接利好大词表语音识别训练;命题小、描述准确。局限是只在单一设定下给出结论,还没有与主流 CTC 实现的实测对比。
Memory-efficient CTC loss with exact loss and first-order gradient equivalence under vocabulary reduction, and activation memory that no longer scales linearly with vocabulary size.
★15