I. Description
1.1 Background Description
Remember in school when the fire is big data, cloud computing and networking, the Internet began after graduating from a lot of discussion of artificial intelligence (as well as block chain). When I heard only when online always feel more distant, I did not mind much attention, until one day forget the afternoon to go back to work or from work in the afternoon, and he was walking on the road when he got artificial intelligence.
He said, you said that we count 1 + 1 is required 1 + 1 is equal to 2, then you know that artificial intelligence is how to count 1 + 1 do, he's not like us so he is counted by learning 0.5 + 0.5 = 1 2 + 2 = 4 so slowly to speculate how much equals 1 + 1, so he calculated 1 + 1 is not necessarily equal to 2, of course, the more accurate the more data for learning results.
When it comes to the little brother also have to mention, although I formed the habit of reading is not affected by the little brother, but the habit of buying books is affected by the little brother of the form, there may be discerned the meaning of point before I thought about going to school or district library, he let me know that you can buy books on Amazon. Of course, recently, a university student's words and deeds let me go to the library feel quite right, so why not start your own should be able to understand actually understand, mainly want to say sometimes the more experience the feeling of Confucius, who say it was very classic, such as the phrase three lines must be my teacher, every person has a unique experience (at least compared to your unique) everyone has their unique experience and insights everyone has their worth you learn from local and even respected. So for now I am always willing to reach out to more people, listen to their stories and ideas. In addition also want to say that I personally do not like have been (or called'm) just do not understand a thing of the parrot remarks about them, even classics such as you do not know what you're against (I think "University" and "moderate" put all pretty good) or even another example, you do not know what Marxism is against you; I quite admire the little brother thing feels he will be able to know a little talking, of course, the most important is their judgment often is not much deviation.
This time the discussion under the influence, but also very high-minded home to pick up just bought near Zhou Zhihua that this "machine learning", and then looked at the complicated mathematical symbols and formulas frustrated to confirm that you do not understand, their machines learning this road came to an end. To tell the truth now about higher mathematics, I probably only vaguely remember differential is derivation, integration is seeking area.
S after the city made a special security efforts, said another formed the habit by buying books to learn, to see Liu Yan's security AI trilogy when looking for a book, but also the AI is safe to buy back. But it feels its content and Zhou Zhihua "machine learning" is very different a bit confused, and then bought a "deep learning and practical framework PyTorch Getting Started" and "TensorFlow Google real deep learning framework", the two front again, and this content another two are not the same, plus the ability and the code was limited, too artificial intelligence can only be considered once again the profound, time is himself a tougher nut to crack.
Also sometimes mention last year and a friend studying artificial intelligence, machine learning, I can only say that an algorithm is not simply a collection of algorithms, but according to their own sense of judgment, these algorithms are not only recent but data a long time ago there was only now able to collect more computing power of computers has come up so fire up the machine learning, although as much as possible but it does not pretend to have a lot of clout confidently.
In fact, often talked about artificial intelligence like this, like a bone stuck in his craw. Pneumonia upset some plan years ago, but the storm at home, but will also have time to win the intellectual framework of artificial intelligence inside the last puzzle.
Of course, there are some twists and turns, one at home a long time work and study efficiency is relatively low, and second, had wanted to make a good middle of the week and smoked a few days to work overtime, the third is connected to GitHub and other foreign sites slower ladder this period has been closed down for a long time even under document fails bad mood, Zhou Zhihua Fourth, on the one hand that this "machine learning" to relatively large shadow on the other hand there are some other things you can do as the siege will face a shortage of attitude a little negative. Luckily, this one is a multi-year program has made great progress, and second, it comes on top of several books can be considered turned over several rounds of heavy head need not understand, the three are the most important read "Web security apparatus yesterday learning portal "this book source code on GitHub each row hundred examples will easily understand.
1.2 Description
This article summarizes the content of large numbers more than Liu Yan from the AI security trilogy: "Web Security machine learning portal", "Web safe depth study and practice", "Web security reinforcement learning and GAN".
Liu Yan GitHub Address: https://github.com/duoergun0729 (three books corresponds to the item followed 1book, 2book, 3book)
In addition to the project directory structure to do some explanation: "Web Security machine learning portal" each chapter and each 1book python file correspondence between the book has clearly stated, "Web safe depth study and practice" is basically a chapter a python code files by a single name correspondence with the file name can be.
Second, artificial intelligence, machine learning and deep learning
The relationship between the three 2.1
Artificial Intelligence> Machine Learning> Deep Learning
Machine learning is the main form of artificial intelligence; depth study is the biggest breakthrough in machine learning forms, but both occur simultaneously when the machine learning generally refers to does not include that portion of the depth of learning.
2.2 Artificial Intelligence
Those who feel there are other branches if-else code can be regarded as artificial intelligence, such as not detected without lights detects the person's door to door, people turn on the lights when this can be considered artificial intelligence.
2.2 Machine Learning
2.2.1 The collection of machine learning algorithms
A collection of machine learning algorithms include: K nearest neighbor, decision tree and random forest algorithm, naive Bayes, logistic regression algorithm, support vector machine algorithm, K-Means and DBSCAN algorithm, Apriori and FP-growth algorithm, implicit Markov Cardiff algorithm.
2.2.2 distinguish each algorithm uses machine learning
There are so many machine learning algorithm, then what use is it to distinguish them, each algorithm for different scenarios do, such as text processing algorithms with this process graphics with another algorithm.
In fact, not a machine learning algorithm is not an individual program these systems an organization put forward, but different people against their machine learning algorithm proposed this concept, these algorithms may be a particular scene this algorithm accuracy higher accuracy of the algorithm is lower, but overall, these algorithms have versatility. Brutalization, that is to place an algorithm can certainly be replaced by other algorithms, such as we are the same can detect brute force, it can also detect WebShell. Or another analogy that we have a lot of sorting algorithms, such as bubble sort, quick sort, heap sort, and so on, all kinds of sorting algorithms speed some differences in the amount of data is not the same situation, but we all can be used to sort.
Nature 2.2.3 Machine Learning
Substantially all of machine learning is the following five steps: the source data, the feature set is determined, the value of the feature set, to learn and use.
Source data: a unified format is a bunch of records; such as Apache logs, database table records.
Determine the feature set: that you have to choose some set of fields, these fields can be set to distinguish between normal and record your goals record. Is high, the rich, handsome is your ideal boyfriend, [high, rich, handsome] on the formation of a feature set (we can also choose additional features to make a more refined judgment).
Feature set numerical: the computer is not good at dealing with "high", "rich", "handsome" operations such language only good at arithmetic value, so in essence we want to achieve on behalf of the numerical values obtained numerical computation. Eg, 0 for not high high 1 represents 0 for 1 on behalf of the rich are not rich, not handsome 0 for 1 for handsome, 0 represents the association represents not worth worth to pay. "High, rich, handsome" -> worthwhile exchanges, it turned into [1,1,1] -> 1.
Learning: We provide [1,1,0] to the machine learning algorithm -> 1 [1,1,1] - a lot of causal data> 0, etc., machine learning algorithms based on these data + code itself sorted out its own algorithm to determine cause and effect mode, this process is learning. (So-called causal judgment is essentially a probabilistic model, which is a lot of people criticized the so-called statistical machine learning, but that is the reason, probability theory)
Use: in the previous step learning machine learning algorithms have been sorted out cause and effect judgment mode, when the input we provide a [0,1,1] when he will be able to give his judgment result (ie 0 or 1) depending on the mode .
2.2.4 Why machine learning is the primary transfer package Man
In the previous section we mentioned machine learning in five steps, that is, when we go to programming codes of these five steps.
The first three steps, and wherein the machine learning nothing and only after the strong relationship between the two-step machine learning relationship, and machine learning algorithm after the two steps of the code itself is fixed and encapsulated in the class class method does not require us written to achieve, at most, we use machine learning algorithm is an instance of the class, to pass its parameters, call its methods.
Or ordering a meaning and we want to do is be ready to sort the array to sort, passed sorting method, and finally to rank. These principles like fast sorting algorithm sorting may be able to write a book, but I am sorry I can not know, I'll just call enough.
Therefore, some of the primary machine learning to see the code, assuming that there are 100 rows, then the data ready line 90 is in the line 10 is in the call class and machine learning. At this level of machine learning programming and general programming a dimensional difference is just that: a machine learning method call is class and machine learning, nothing more.
In addition, "10 lines is calling machine learning class and its methods" also means that we replace from one to another machine learning method machine learning method, just put the "10 line" from "calling this machine learning algorithm corresponding to the class and its methods "to" call that class and its methods of machine learning algorithms corresponding ", and nothing more.
Of course, we want well-founded, "Web security machine learning portal" Many examples are exactly the same, such as: https://github.com/duoergun0729/1book/blob/master/code/5-3.py
2.3 depth study
2.3.1 depth set of learning algorithms
Depth learning algorithm set includes: convolution neural network algorithm (Convolution Neural Network, CNN), recurrent neural network algorithm (Recurrent Nerual Network, RNN).
2.3.2 The difference between deep learning and machine learning
Briefly, machine learning is to "a given predetermined algorithm code data +" to generate a feature for each of the specific gravity and the relationship of correspondence between the results. For example, in the above example exchanges, it is assumed come to your liking each feature a high proportion relationship is ultimately the proportion of 30% to 30% the proportion of point-rich handsome accounted for 40%, assuming that the outcome of the relationship is based on your preferences of a specific if the individual is calculated to give less than 0.6 is not recommended to exchange greater than 0 0.6 1 recommendation is given exchanges.
To say the depth of learning and machine learning essential difference, I'm not quite sure, we can only summarize a number of statements about the use of the Internet. First, the data needs of machine learning can be less deep learning data have a little more advantage, the second is better at CNN in image recognition, RNN has more adept in terms of the timing of the operation.
2.3.3 When we use deep learning the code is a little more complicated than learning machine but also a little bit
In 2.2.4 we say that the use of each machine learning algorithm is relatively simple, it is not the depth learning to use it will be a little complicated.
First, the answer is yes. But we must be clear that we want to believe that is still the same algorithm are established that we do not need to be written, five steps are the same, deep learning of complex call does not come from the complex on his principles, but only from we can set the parameters a little bit more.
Third, machine learning framework
In fact, first, because of machine learning is not a relationship between one independent algorithm composed of two machine learning the code is relatively short, it is generally not how to speak frame. He said frames are generally deep learning framework, such as AlphaGo is based TensorFlow and thunder are increasingly Sheng PyTorch deep learning framework.