Types of neural networks
performer desp_hero
Artificial neural networks are the most promising technology in the field of artificial intelligence at the moment. They owe their name to the fact that they seek to mimic the neural network of the human brain at the physical or software level. It is believed that it is thanks to the unique structure of the neural network in the human brain (and most mammals, reptiles and birds) that the ability to learn appears. This is exactly what the basic idea of artificial neural networks is based on — "What if programs can be trained instead of written?" Accordingly, the main classification of neural networks is related to their training methods.
There are only two main teaching methods — with a teacher and without a teacher. Learning with a teacher implies that there are two vectors — input and output, and, using trial and error, as well as the desire to reduce error, the neural network strives in one way or another to reduce the output vector to the input, after which all positive changes are fixed and negative ones are removed, and a new input is stored in the network memory. vector.
Teaching without a teacher works much more elegantly and is considered a more progressive method of teaching. To explain it in simple terms, the network does not have an unambiguously correct answer to the questions asked (input vectors), after which it would be possible to move on to the next question. In other words, after the network's response, it is not known whether it answered the question correctly or not. The learning process consists in the fact that the questions (input vectors) do not differ much from each other, and, accordingly, the answers to them should not differ much either. Thus, the network literally learns by itself, trying to minimize the spread of answers to similar questions.
As for other classifications, networks are often classified by the type of connections between neurons — by the number of input and output connections, as well as by their dynamic or static nature — in other words, whether connections change over time or not. The most popular networks at the moment use static connections — the number of output and exit connections for each neuron is precisely defined and does not change, just like the neurons to which these connections lead. Also, do not forget that most software neural networks operate exclusively with binary logic, while some physical neural networks may use analog or inaccurate logic. Descubre el increíble juego de Burger Win sin depósito inicial