Reporting by XinZhiyuan
Editor: Good sleepy Yuan Xie
In order to smoothly refine the model, friends who engage in AI have really thought of any way.
Machine learning, commonly known as "alchemy".
As a kind of AI "black magic", what about adding a little "metaphysics"?
Recently, a PhD student from Soochow University did a "random collision" project.
The old brother said that the random seed to be used in machine learning will affect the final experimental results, so it is better to do a blessing of luck.
Open Source Projects: https://github.com/Spico197/random-luck
This is really "the psychology of the East China Sea and the West Sea is the same, and the South Learning and the North Learning Taoism have not cracked."
There are two ways to generate numbers: one is to simply fill in a birthday date, and the other is to calculate the number of the sky and dry branch of the AI experiment start time, return a lucky number, and use this number as a random seed to run the experiment.
The effect goes something like this:
The code is as follows:
Although the author says it's just for fun, he doesn't want to attach a scientific proof or explanation.
However, the code farmers and students who are widely suffering from the transfer of ginseng can also pin some good wishes on this... Right?
The author, Zhu Tong, is currently pursuing a Ph.D. from Soochow University, after receiving a bachelor's degree from Guizhou University.
His research interests include IoT design, relation extraction and event extraction.
He won the first place in the CCKS "Event Extraction of Financial Documents" task team in 2020, the third prize in the Language and Intelligence Competition (top 5 teams) in 2021, and the CCKS Event and Relationship Extraction Task Technology Innovation Award.
Heaven and earth branches
Although few of the code farmers with a scientific spirit now really believe that the dry branch is really a real commodity, it does not mean that the people of the times who are keen to use the concept of dry branch also see it this way.
According to the research of Guo Moruo, the great master of the ancient script examination circle, the "Twelve Branches" and "Tai Nian Chronology" that first appeared in ancient China in the 13th and 12th centuries BC originated from the "Zodiac" system that first appeared in the 45th-23rd centuries BC and matured in the 13th century BC.
The twelve "Tai Nian Nian Names" corresponding to the "Twelve Branches" are transliterations of the ancient Sumerian and Akkadian words of the "Zodiac". The constellation symbol of the "Zodiac" on the clay tablets of ancient Asia is also almost identical to the oracle bone image of the "Twelve Branches of the Earth" later.
Chart of the constellations of Ancient Asia
That is to say, the fortune-telling technology of the Celestial Stem-Earth Branch System and the Twelve Constellation System actually has a similar origin. Just like CNN now knows no borders, everyone has to roll a volume...
And the ancient West Asians and ancient Chinese thousands of years ago looked at the Star Palace/Ganzhi, and the contemporary AI practitioners looked at LeCun: they are all cutting-edge technology, even if I don't understand, but the big bull must be very powerful.
Therefore, taking the simulator to spit out a dry expenditure is a combination of ancient frontier technology and contemporary cutting-edge technology.
After all, the science fiction novelist Hein rhine once said: "The supernatural phenomena that you think are so powerful are all precision engineering phenomena that the creators find troublesome."
Random seeds
Conceptually, seed values are used to generate random number generators. Also, every time you use the same seed value, you get the same random value.
That is, a random seed ensures that anyone who reruns this code will get the exact same output.
Typically used in two tasks:
1. Split data into training/validation/test sets: Random seed ensures that data is split in the same way each time you run your code
2. Model training: Algorithms such as random forest and gradient boosting are non-deterministic (the outputs are not always the same for a given input), so a random seed parameter is required to obtain repeatable results
In addition to reproducibility, random seeds are also important for referenced results. For example, when testing multiple versions of an algorithm, it is important that all versions use the same data and are as similar as possible (except for the parameters that need to be tested).
Some people use the same seed every time, while others generate it randomly.
For example, when training a machine learning model using Scikit-learn, a function imported from a sklearn.model_selection module train_test_split use parameters such as random states to get the input to a random seed.
Random_state=42 this parameter sets the random seed to the same value each time the above code is run, and the same set of validations (X_test, y_test) is obtained.
AI write your daily astrolabe
Of course, such a romantic idea of AI demonization will not be limited to Chinese code farmers. The idea of writing an astrolabe with AI has also been used by American code farmers.
In 2017, VICE magazine reported that three New York code farmers with code power, product development power, and horoscope hobbies, suffered from the "smart astrolabe" and "automatic astrolabe" app on the market at that time, which only re-read the works of Susan Millers in the 1990s, and made a "astrolabe algorithm" themselves.
The basic mechanism is not complicated. The developers first entered the corresponding fortune-telling interpretations of various celestial movements in astrology, and at the same time developed an interface to automatically extract the celestial tracking data published by NASA. Algorithms are then developed to map the calculations of this data, and natural language generators are developed so that the algorithm results appear as "Your Daily Astrolabe" article.
These horoscope enthusiasts also have very qualified product manager skills, and the "Astrolabe Algorithm" app places special emphasis on the social function of adding friends. Not only can you compare each person's astrolabe table, but you can also generate a stellar fit between each person.
The "entering the palace" of the interaction between various celestial bodies and zodiac signs, the influence and interaction of the fate of each person and each person's friends, is also presented on the app. The more friends you add, the more specific the results will be presented.
The developers of the "astrolabe algorithm" happily said to VICE, "Of course we know that this is not scientific fortune telling. But it's an interesting way of telling a story, based on both ancient mythology and contemporary personality, and it's fun."
But not only is romance universal, but the personality of science students who rely on strength and singleness seems to be universal. In 2020, there were code farmers who went to Twitter to sample the emoji frequency of astrolabe Twitter replies as raw data, and wrote a project to run statistics. Take a stance on the accuracy of the respondent to verify whether astrology can really be effective in fortune telling...
Nonsense, of course not, really do not understand the style, do this project people will not have a girlfriend.
Resources:
https://github.com/Spico197/random-luck