推荐nl · 2015-03
“It's a pretty revolutionary approach - no one seems to have tried anything very similar since the 1990's.”
推荐它,称这是相当革命性的做法,自 1990 年代以来无人尝试过类似方法
给神经网络配上可读写的长期记忆来做推理与问答
论文提出一类称为记忆网络(memory networks)的新模型:把推理组件与可读写的长期记忆结合起来,让模型学会联合使用它们。在问答场景中,长期记忆相当于动态知识库,输出为文本回答。作者在一个大规模问答任务和一个由模拟世界生成的小而复杂的玩具任务上评估,后者展示了模型串联多条支持句来回答需要理解动词意图的问题的推理能力。
评论者把它看作早期把可读写记忆接到神经网络上做推理的新颖思路,常与 Jason Weston 的系列工作及后来的 DNC 一起提起,也有人把它回顾为类似 ChatGPT 的早期能力雏形。
“It's a pretty revolutionary approach - no one seems to have tried anything very similar since the 1990's.”
“This would also fit well into memory networks: Memory Networks, SDM: [LINK]”
“Jason Weston (Facebook) etc's work on Memory Nets[1]”
“There's been a strand of research mostly coming out of Jason Weston's Memory Networks research[1].”
“So we can see quite early ChatGPT like capabilities”