Recently, zhejiang’s college entrance examination essay “Living on a tree” went viral on social media because of its obscure content. The layman sees the fun, the layman sees the door, the gods see the essence through the appearance.
Generally speaking, the composition of college entrance examination can achieve a satisfactory score through mass and patterned training. In fact, it is to train students as a writing machine, providing students with a large number of model essays, writing patterns, and then a long period of study. Wait, isn’t that the same training process for artificial intelligence?
So if you give an AI system the same data, can you teach it to write a composition?
The answer is yes, of course.
01 Basic Components of EssayKiller
The core of an AI, regardless of its shape and function, is an artificial neural network. EssayKiller is also composed of four different neural networks.
The first part is the recognition network similar to the human visual processing system. Through real-time OCR of external camera, convolutional neural network and recognition and extraction of Chinese characters, it inputs the results to the network of the next layer.
The second part is the language network, which aims to model the language function of the human brain. Here, the UP leader divides it into two sub-networks, which are more concise and concise compared with the complex structure of human temporal lobe and frontal lobe. One subnetwork is the reading topic network, which is to accurately read and extract the topic summary of the college entrance examination.
Another subnetwork is the writing network, that is, the writing of articles based on topic summaries. The latter is the core of the whole AI, and EssayKiller is able to understand and output Chinese characters based on this neural network.
The third part is the classification network. EssayKiller takes just 0.1 minutes to write an 800-word gaokao essay, which normally takes 40 minutes for a normal student. Apparently, EssayKiller can take advantage of its ability to write quickly, generate multiple essays in its brain at the same time, and then choose the best one to answer in the same time.
The idea borrows from AlphaGo’s design. He used the API to invoke a separate neural network to score 100 essays for smoothness and output the essay with the highest score. In theory, high scores could be achieved easily by knowing the preferences of the test makers and making the AI perfectly fit the essay that the markers liked.
This is similar to the process of training students to write essays for the college entrance exam, which has been used in many recent years for full marks, such as “Living on a tree”.
Code words are not easy to nonsense two sentences: need python learning materials or technical questions to exchange “click”
02 Learning process of EssayKiller
It sounds simple enough, but EssayKiller was born without the four network building processes that the gods would bow down to as soon as they operate it. At this point it is still a blank sheet of paper and needs to be fed a lot of data. As the saying goes, only by reading ten thousand books can YOU write a good composition.
And then EssayKiller is not a long learning curve. Give it a large number of high-quality prose, argumentative essays, such as modern essays, college entrance examination compositions, and so on, and then let it self-training and improve writing ability. Out of personal preference, prose writers such as Lin Yutang, Mu Xin and Lu Xun were added, as well as some modern writers such as Wang Xiaobo, Shi Tiesheng and Wang Shuo.
Finally, with 3.6 million images, 200 million Chinese pre-training materials and thousands of fine-tuning articles, EssayKiller started its “long” journey of learning with the love of UP.
As it turns out, no one can casually succeed. At this point, there was a “happy” episode. OOM, which stands for “Out of Memory”, translates as “Out of Memory” in Chinese.
EssayKiller’s neural network has reached a staggering 1.7 billion inputs, and the average graphics card can no longer support such a large network size and computation. After several failed attempts, EssayKiller surpassed the computational memory limit of any single GPU currently on the market. Finally, the RTX8000 was chosen with huge investment and EssayKiller started training successfully. Sure enough, raising kids costs money.
After 137 hours of non-stop learning, EssayKiller finally became a qualified high school student. Now it’s time for an exciting test. Choose two national exams and two regional exams. EssayKiller uses its multi-module heterogeneous deep neural network with 1.7 billion neural network parameters, which EssayKiller is proud of, to output the text in real time and transfer the essay to the answer sheet within the specified time through the typeset script written for the college entrance examination format, combined with the modified external device. Finally, I contributed four very good compositions for the college entrance examination. (The writing process is like the giddy GIfs at the beginning.)
Xiaobian selected the composition results of Zhejiang paper to show the writing strength of EssayKiller.
(Screenshots have been trimmed down for easy reading. The right side of the article shall prevail.)
As a Python developer, I spent three days to compile a set of Python learning tutorials, from the most basic Python scripts to Web development, crawlers, data analysis, data visualization, machine learning, etc. These materials can be “clicked” by the friends who want them