CS224D Lecture 12 Notes

Welcome reprint, Reprint annotated Source:

Http://www.cnblogs.com/NeighborhoodGuo/p/4711678.html

CS224D's 11th lesson is their class exams, so the video on the back jumps straight to lecture 12. Lecture 12 is a foreign guest who is said to be a Facebook engineer, speaking of the most advanced AI Research,ai there is no solution in the world or a unified view of a problem, so this class is not to introduce a mature model, It is understandable that this is to open a door for everyone to understand how this part of the AI work is carried out, from the perspective of very super-super-high-level explained this.

This class introduces a thing, is the memory Networks. But the memory networks covers a wide range, plus the model is immature, and a lesson is just a brief introduction to it.

Memory networks is not the kind of memory that we have defined in the model we talked about before. The memory in the model that was introduced in the previous lesson is the use of various data and then training model, you can think of memory stored in the model weight, can be used for object detection and other work. And this lesson is not a concept of memory and previously defined, the memory similar to the computer software, this class to memory similar to the computer's hard disk, need to use memory to extract data from the hard disk, when not needed to put there.

The AI in this class is just a small branch of AI, which is part of the Q&a, which is similar to the reading comprehension of our middle school English test, or the last few reading questions of Chinese college entrance examination. The goal we actually want to achieve is to allow the computer to complete the level of the reading question in the Chinese college Entrance examination, which is a good idea, but it is difficult to measure the good and bad of the existing model, and it is impossible to implement such a complex architecture. So, based on the above considerations, the intelligent scientist divides the q&a into several pieces of English reading comprehension which is similar to our middle school, does not need to answer the complete sentence, as long as the English reading like the division Select one or just answer single or several key words. There are specific explanations and examples of tasks that are presented in the paper, from the previous to the post-sorting is easy and difficult. We generally think that after the completion of the above several pieces of the task, and then use an optimized model so that the output of the sentence comparison nature can complete the q&a task of AI.

The goal of this memory network is to be able to understand (comprehension) an article or a movie and then answer questions based on the content of the article or movie. But there are so many kinds of verbs, if a word is not realistic, the efficiency is very low, and the meaning of many verbs is actually similar, so you can use the basic command to replace these various words.

When the substitution is complete, the process is relatively straightforward.

With the processed data, there are a number of simple q&a targets that can then be set up in the model.

The model we set up is divided into four parts i,g,o,r

I this module, in fact, most of the previous courses are for the tasks contained in the I module, such as PARSING,RNN.

The search vectors for the O module, whether in the paper or in class, are searched for only two, as follows:

The simplest is to take out the highest score vector at a time, and then generate seemingly nature answers based on the two vectors and the original problem.

What if a word has not been seen in memory or a word is lost? The easy way is to use the nearest word and then Bag-of-words

Of course, the above content is far from the requirements of AI, in class and the paper gives the approach AI two methods:

OK, the gate of AI has been opened! I wish I could see the realization of AI in my lifetime.

CS224D Lecture 12 Notes