Over the past few years, there has been a significant increase in the performance of artificial intelligence systems. This led to the fact that every second company now declares the use of AI in its activities. Despite the rapid development of this technology, it is still inherent in serious disadvantages that are considered in the material.
Now the study and use of artificial intelligence in most cases is reduced to the creation of a neural network, the algorithms of which appear as a result of machine learning. The problem is that the purpose of learning reduces to the conclusion of the most effective work algorithm for a particular task. This is not quite suitable for the concept of artificial intelligence. AI implies a robotic brain that thinks just like a person — about everything. A more accurate name of such a mechanism — AGI (artificial common intelligence). He thinks about solving different tasks. Machines capable of emulating human consciousness completely, does not yet exist and even
Based on statistics of posts of mail letters Gmail AI offered to answer the same type of letters «I love you.» This phrase approached statistics, but I could not understand that a similar answer to the boss could lead to dismissal. This happened due to the fact that the essence of the phrase was not taken into account at all or the learning algorithm turned out to be incorrect. On the way to acquire consciousness, the disadvantages of artificial intelligence systems were often manifested in amusing cases like sexism Ai Amazon or hazardous recommendations for sick cancer from IBM Watson. Cons Ei also managed to manifest themselves in games:
An experiment was even more fun, in which the car gave complete freedom and freed from any purpose for the day. It was assumed that the AI would be improved, but the computer watched TV all day, diluting this occupation with games in Super Mario Bros and the classic from Atari 2600.
A visual example of the operation of the neural network — recognition of images. If you put the task of determining the presence of a dog in the picture, then to neurallet will learn from each photo. During learning, all images are marked (there is a dog or not). Each photo of the neural network itself assigns his probability ratio of the presence of a dog silhouette. The machine learning process consists in correcting errors. When there are no dogs in the photo, and the AI finds it, an error is issued. Then the necessary corrections and repeated tests occur. All these actions occur without human participation. As a result of a certain number of iterations, neurallet is trained with a high probability of properly determining the dog in the picture.
The problem is that dogs and cats it does not distinguish itself — this is another parallel task. It is necessary to solve it with the introduction of new data structures for cats and holding a new learning stage. Any other animal will also require a repetition of these actions in a neural network.
This type of neural network training is divided into 4 subspecies:
The first method is considered the closest to the ideal artificial intelligence, since the testing environment is acting as a teacher for a neural network or even another neural network. The most successful example of depth education with reinforcement — chess neural
If you look at the training of Parkour, the results will be no longer inspiring. In the role of the educational environment, all items with which the tracer is faced on the way. Training with reinforcement as much as possible to real intelligence capable of processing data even about an unfamiliar environment. The reverse side of this universal method is a large amount of processor time spent. Deepmind has to train the tracer Parkour with Deep Learning left as much as 6400 hours, while Boston Dynamics developers were able to embody all the necessary algorithms in a real robot completely without deep learning. The results are impressive.
Conclusion — deep learning in the AI can sometimes be successfully replaced by the long-proven classical methods and algorithms. Until the universal general artificial intelligence has been created, some tasks can be successfully solved without it.
The leaders of various projects and startups seek to use new-fashioned technology everywhere, sometimes just for
Create a list of the best customers of the week, identify users who have not made purchases for a long time, contact customers by name, increase the consumer loyalty index — all this can be ordinary SQL requests, write and implement that are not so expensive. With artificial intelligence, this work will cost much more expensive due to a higher salary programmers.
Immediately arises a logical question: «What is the advantages of using artificial intelligence in online stores?». We give a simple example.
Man bought sneakers for running. For targeted advertising on SQL requests, you need to create an additional database for goods that can now be useful to this buyer (T-shirts, shorts, socks for running, etc.). Consider that the goods are different and connections between them a lot. In addition, some products refer to the category of fast spheres (socks), others can be changed for years (T-shirts). This means that some purchased things can be offered again, others — no. Even connections themselves can be both direct and indirect. It is in processing such data with different links between things that the best results show ML / AI. Recommetive goods generated in advertising from AI falls in the top ten much more often than goods from advertising using SQL queries.
Summing up, it should be noted that AI is still very far from acquiring consciousness. Perhaps it will never happen, but most of the existing minuses of artificial intelligence after a few years