[Poll] And that's about neural networks, AI, etc.

On the Internet, it has often been seen that the de "neural network is a panacea for everything and everyone," that is, for example, “set a neural network - and everything, profit” or even more brutal “will soon create an AI based on a neural network that can replace even programmers / administrators / analysts, etc.”.

As a rule, when you try to figure out the depth of knowledge of the Zen opponent in this area, it turns out that he is almost a complete amateur (or, after a couple of stupid "arguments" with foaming at the mouth, he merges, thereby interrupting the attempt to find out his level).

I am familiar with neural networks, participated in some projects, "designed", trained and configured more than a dozen different flows using the latest in various applications (while having tried many different engines and types of networks from simplest perceptrons to self-organizing monsters). In addition, I am personally acquainted with a number of people actively designing and using neural networks in everyday life, and so far have not heard from any of them about the stunning breakthrough in the use of neural networks (with regard to the latter’s intelligence). Those. I think that I can voice the distance from the reality of the above statements.

I wanted to write an article (and a poll) for a long time, but somehow I didn’t reach it. And after the next question-proposal on the e-mail "set the neural network on the same way" to the problem from my previous article "Monitoring log-journals: Such a vulnerable log ..." , nevertheless I understood - no, I need to write.

As a result, the purpose of this article I put firstly to express my modest opinion, and secondly to find out what professionals think about this. Those. so that you can poke your nose into the article of the next such "all-better-all-knowing" from the Internet ...

Although all of a sudden I was still mercilessly behind the times ...

So, to begin with, voiced about or rather the contra of the use of neural networks, as for its “intelligence”:

• the neural network is “programmed” to solve narrowly focused tasks, i.e. the network disentangles only strictly defined, semantically significant features for it;

  
  

• the ability of a neural network to disentangle features of the input data directly depends on the quality of the input data (for example, to what extent the semantically significant features fit into the network “creator” picture); for example, it does not distinguish a cucumber from a banana if it does not initially contain the function of assessing the color component or a number of other characteristic features (such as it will find a banana only in the monkey’s hands, since the latter will hardly find a cucumber somewhere ) or they are simply absent in the input data, and additional “knowledge” was not laid by the network creator

  
  

• The data from the training sample on which the network itself is trained should be:

  
  
  • are representative (must illustrate the true state of affairs in the subject area);
  • consistent (inconsistent data will lead to poor quality network training);
  • transformed to a form that can be sent to input neurons (i.e. must be at least primarily processed by other algorithms)
  
  

• the results of the decision after the training and in time must be checked for the correctness of the decision (who will do this when we are all sent to retire ...)

  
  
• complex neural networks for the most part are very, very slow creatures that also consume a lot of resources and memory (for example, to store the same associative information), although solving some narrowly focused, pre-set tasks is often faster than a person (well, a calculator can, for example, build to a degree millions of times faster than a man, that’s why he hardly became smarter than the latter and boasts greater intelligence).

Roughly speaking, most of the classic neural networks that exist at the moment are classifying functions or groups of functions that more or less efficiently approximate input data to the required limits and, as a result, produce strictly strictly programmed features.

Those. not some abstract thinking "glands", but the essence of specialized algorithms, strictly or narrowly sharpened on certain actions, with the ability to "self-study" (in fact, independently improve or refine the parameters of the "algorithm" in the learning process, and the networks are trained without teachers, with a permanently on training mode, also in the process of work).

Those. as far as intelligence is concerned, the fundamental difference is how a neural network differs from any algorithm (and this is one of its main advantages) - it is able to improve the result based on past experience (I exaggerate, but this is intelligence, we’re not talking about connectivity here , neurons, signals, weights, etc.).

In addition, often the result of the work of such an "algorithm" is a probabilistic estimate (or a group of them). What I want to say by this. Well, take for example the log line from the above post :

Aug 18 08:04:51 srv sshd[2131]: Failed password for invalid user test from 1.2.3.4 port 46589 ssh2 from 4.3.2.1 port 58946 ssh2

" 1.2.3.4 99%", , " from " ( ). — "" ( flow ).

, " " (IDS), ( ) , ( ), , . "".

"" (.. ) — , ( , ) .

, , .

, - - siri , "", , .. - ( ). , .

" " — , , .

, - -, :

• , , .. , , - ..
• ( )
• ( )
• ( -)

.

"" , ( , — ), :

• , ( // ).

  
  

• , .. , , ( )

  
  
• : ( -).

, ( "" )...

, !