|
With the invention of neural networks and methods associated with them, machine learning has become possible, the character of which is of an accidental nature. ie People can not accurately predict exactly how and when the neural network will learn something.
Random search for successful knowledge
At its core neural networks are trained in a full search of all possible options, until the successful options are met.
This method is very long and resource-intensive, but in the absence of any additional support points the only possible one.
The machine with which Alan Turing deciphered the German Enigma system by a complete search and listened to the whole radio exchange of the Germans during the Second World War
Let us give an example. We will consider the primitive man as a neural network and see how many years have passed for the emergence of certain technologies by this method of training.
0. We shall consider conditionally that a human has arisen 10 million years before our era
The primitive man in the artist's view
1. Stone ax 3.3 million years BC
2. People took fire 1.5 million years BC
3. Writing 6 500 years BC
4. Wheel 5000 years BC
5. Ships 4000 years BC
6. Money 3000 years BC
7. Pyramids 2500 years BC
8. Iron 1200 years BC
9. Proof Geometry 500 years BC
10. Algebra 820 year AD
11. The law of universal gravitation 1666 AD
12. Differential calculus 1682 year AD
We see that for the first discoveries - a stone ax, fire, writing, a wheel - an insane number of random attempts to train a neural network took place, millions of years were passed since the appearance of man.
But further, relying on these basic discoveries, things went much faster: on the basis of the first group of discoveries, ships were built, metals were smelted, pyramids were built.
The level of civilization of the second group of discoveries made it possible to invent a systematic mathematical and physical science.
The level of third discoveries gave rise to a modern technological civilization. ie When inventing artificial intelligence there are 2 ways:
1) evolutionary - launch a very powerful neural network and just sit and wait for a couple of million years, until she at least learns something
2) revolutionary - to teach the neural network all the most important thing that is known to modern science, so that it will be based on something and go much faster.
|
