Paper reading: Memory Networks

I. Problems solved by the thesisThe realization of long-term memory (a large number of memories), and how to read and write from long-term memory, in addition to increase the reasoning function why long-term memory is very important: because the traditional RNN even replication tasks are not, lstm is expected to be enough iffy.

In QA problems, long-term memory is very important and serves as a knowledge base. The problem with getting long-term memories from them is that when there are several sentences and there is a connection between the sentences, the RNN and lstm can not be solved very well and assume a long-term dependence. Need to extract information from memory second, the solution of the thesis (0) Overview of the overall architecture the so-called memory network is a common framework, the internal input mapping, update memory mapping, output mapping, response mapping are replaceable

Watermark/2/text/ahr0cdovl2jsb2cuy3nkbi5uzxqv/font/5a6l5l2t/fontsize/400/fill/i0jbqkfcma==/dissolve/70/gravity /center ">
(1) A natural language-specific implementation of the I module takes an input text. The text is stored in the next available memory slots in its original Formthe G module are thus only used to store this new Memory, so-old memories is not updated.

Watermark/2/text/ahr0cdovl2jsb2cuy3nkbi5uzxqv/font/5a6l5l2t/fontsize/400/fill/i0jbqkfcma==/dissolve/70/gravity /center ">



Thirdly, the scheme in the paper solves this problem. How far has it been resolved?
In fact, the main paper is to mention a large framework, and the framework of each module is able to change, so that can adapt to different applications, but in fact, there is no significant innovation. is essentially a RNN type of network.

Iv. Experimental (1) Experimental results

Paper reading: Memory Networks