循环神经网络:原理、架构与应用
1. 自回归与文本生成
自回归在文本生成中展现出了独特的优势,它可以从源材料的随机位置开始生成连续字符,并且能够根据需求持续运行,生成无限量的输出。在训练的第一个周期后,我们可以看到初步的生成结果:
er price.” “If he waits a little longer wew fet ius ofuthe henss lollinod fo snof
thasle, anwt wh alm mo gparg lests and and metd tingen, at uf tor alkibto-
Panurs the titningly ad saind soot on ourne” Fy til, Min, bals’ thid the
虽然此时生成的“单词”并非真实词汇,但它们的长度与英语单词相近,且并非随机字符的组合,很多甚至易于发音。经过50个周期的训练后,生成结果有了显著改善:
nt blood to the face, and no man could hardly question off his pockets of
trainer, that name to say, yisligman, and to say I am two out of them, with a
second. “I conturred these cause they not you means to know hurried at your
little plat