Content




Different forms of artificial intelligence

I did not mean semantic text search, sentiment analysis, image recognition or other technologies used for the tasks not reduced to regression or classification. Several times during my online searches I ran on pages explaining new training methods for robotic devices. These pages later disappeared without traces. From what I've seen and read, I can conclude that neural network hit the dead end in development and will be replaced soon by something drastically new.

The learning system will be designed to recognize the goal by observing another robot or working human. Besides the goal recognition, it will be designed to identify obstructions and rules to follow. Without breaking the rules AI should be able to find the room for optimization and outperform the observed training subject. This AI suppose to be built on other than neural network basis, but may use it as an internal element.

For example, a driverless car. The designed on above principles AI will be able to recognize goal to ride from A to B, learn traffic rules, start recognizing obstructions on the roads and optimize routing and speed simply by observing actions of human driver and environment.

It is clear from some indirect evidence that big corporations are quietly moving into this direction and it is possible that very soon experts in neural networks will find their skills no longer demanded.

But this site is not innovative to such degree. Here we promote one very effective model which shows the same accuracy as neural networks but outperforms them in learning rate. It was introduced in 1957 by two Russian mathematicians Andrei Kolmogorov (first from the top) and Vladimir Arnold (second from the top) $$ M(x_1, x_2, x_3, ... , x_n) = \sum_{q=0}^{2n} \Phi_q\left(\sum_{p=1}^{n} \phi_{q,p}(x_{p})\right). $$ The right hand side of above equality is an hierarchical tree of discrete Urysohn operators, introduced by Russian mathematician Pavel Urysohn (third from the top) in 1920th, which continuous form is below: $$ \phi ( x) = \lambda \int\limits _ \Omega K ( x , s , \phi ( s) ) d s + f ( x) ,\ \ x \in \Omega . $$ The numerical algorithm, used in training and showing an extreamly quick covergence rate, is slightly modified method of projection descent, published by Polish mathematician Stefan Kaczmarz (fourth from the top) in 1937.

All novel ideas and methods are published but the site provides explanation in reader's friendly format.