arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1702.06186cs.LGcs.AIcs.NE
收藏

神经网络推理综述

Survey of reasoning using Neural networks

TU Kaiserslautern(凯泽斯劳滕大学)

Amit Sahu

摘要:

推理和推理需要人类的过程和记忆技能。神经网络能够处理像图像识别这样的任务(比人类更好),但在记忆方面仍然有限(受注意力机制,大小)。递归神经网络(RNN)及其修改版本LSTM能够解决小的内存上下文,但当上下文变得大于阈值时,很难使用它们。解决方案是使用大的外部存储器。尽管如此,它提出了许多挑战,如,如何训练神经网络的离散内存表示,如何描述长期的依赖关系,在顺序数据等最突出的神经架构,这些任务是内存网络:推理组件结合长期记忆和神经图灵机:神经网络使用外部存储器资源。此外,需要额外的技术,如注意力机制,离散存储器表示上的端到端梯度下降来支持这些解决方案。上述神经架构在简单算法(排序,复制)和问答(基于故事,对话)应用上的初步结果与最先进的水平相当。在本文中,我解释了这些架构(一般),使用的附加技术及其应用结果。

英文摘要:

Reason and inference require process as well as memory skills by humans. Neural networks are able to process tasks like image recognition (better than humans) but in memory aspects are still limited (by attention mechanism, size). Recurrent Neural Network (RNN) and it's modified version LSTM are able to solve small memory contexts, but as context becomes larger than a threshold, it is difficult to use them. The Solution is to use large external memory. Still, it poses many challenges like, how to train neural networks for discrete memory representation, how to describe long term dependencies in sequential data etc. Most prominent neural architectures for such tasks are Memory networks: inference components combined with long term memory and Neural Turing Machines: neural networks using external memory resources. Also, additional techniques like attention mechanism, end to end gradient descent on discrete memory representation are needed to support these solutions. Preliminary results of above neural architectures on simple algorithms (sorting, copying) and Question Answering (based on story, dialogs) application are comparable with the state of the art. In this paper, I explain these architectures (in general), the additional techniques used and the results of their application.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑