A model of self-modificated predicate network with cells implementing predicate formulas in the form of elementary conjunction is suggested. Unlike a classical neuron network the proposed model has two blocks: a training block and a recognition block. If a recognition block has a mistake then the control is transfered to a training block. Always after a training block implementation the configuration of a recognition block is changed. The base of the proposed logic-predicate network is a logic-objective approach to AI problems solving and level description of classes as well as the notion of partial deducibility which allows to extract common sub-formulas of elementary conjunctions